Hello everyone, this is Karthik, the previous president of NYP AI. The time has come for me to write this final blog post. This marks the last blog from the current NYP AI Team, soon to be succeeded by Wen Bing and his team. This blog aims to remind our current excos and members what events we have worked on and also share some of the lessons we have learnt along the way so that the next team can set the bar even higher.

Highlights of the Year 🤩

So without further ado, let's talk about the 3 of our biggest events of the year:

PentaHacks 2023

This was the first event where we collaborated with the AI clubs from other schools such as Singapore Polytechnic, Ngee Ann Polytechnic, and more. We organised a hackathon in early March of 2023 and saw a very good turnout of over 200 participants. This first event was invaluable because it helped us expand our connection to other schools which will open up more doors for collaboration. It also showed us that we needed to partner with other organizations like AI Singapore which could help us with some of the resources which we needed to fund some of our events as well as help us to reach out to the larger student network and give them resources to learn more about AI.

AWS Deeprace Dash Hackathon 2023

This was the second event that we did, in the middle of July. AWS was actively promoting students to learn AI skills and they were also hosting this competition called the AWS nationwide deep racer hackathon. To help give students some hands-on experience for the hackathon they reached out to some AI interest groups in Singapore to get us to help out for the event. The event saw over 400 participants and was open to all students from secondary school, junior colleges and polytechnic and was held at the AI Singapore office. We also had the opportunity to showcase some of our club's projects, which helped increase awareness and interest in our club.

Panel Discussion with AI experts

Just a week after the Deep Racer hackathon, we hosted our AI panel discussion, our largest event to date, with over 500 participants. Thanks to AI Singapore and Ms. Teo Miow Ting, who helped connect us with industry professionals for the panel. The discussion covered topics like generative AI, opportunities for students in the field, and how to start learning about AI. We received excellent feedback from participants, indicating the event's success and impact.

Lessons Learnt 🤔

While we hosted many successful events, we also did have a few which were less successful. We found that our highly technical workshops attracted fewer participants compared to our beginner-friendly ones, despite covering cooler topics. As a result, we decided to focus on larger-scale events like those mentioned earlier which catered to all skill levels from novice to pro.

We also learned the importance of effective publicity. To improve this, we started putting out short Instagram posts explaining AI concepts in a simple manner, created a Linktree to share all our resources better (link here: https://linktr.ee/NYPAI), and even came out with our own T-shirt. Recognizing the need for a stronger brand identity, we’ve put more effort into marketing our workshops and events, ensuring we reach a broader audience and make a bigger impact.

Epilogue

All this stuff wouldn't have been possible without the NYP AI crew (Wei Jun, Jay, Jun Ming and Jing Khai) who worked tirelessly to make this happen. We are proud of what have built so far and how we were able to grow this community for AI enthusiasts to share new ideas and collaborate.

All the best to the next team, we hope you continue to grow this community to be even bigger than it currently is!

~ Sincerely
The NYP AI Team (23-24)

Join Us | Website | Instagram | Discord | Telegram

We forayed into unknown territory from the get go this year.

Late April, we hosted our first ever event engaging our club members in-person this year in the large space at P1 Convention Centre in school. Titled An Enchanting Escapade by NYP AI, the aim was to educate students about AI concepts through games while promoting our club's awareness.

Time – the biggest enemy.

The event comes after the launch of our new Community Engagement (CE) Department within our EXCO, spearheaded by Shawn Goh (Year 2 in Diploma in Applied AI and Analytics). The lack of precedence, since an event of such a scale has never been done before by any Student Interest Group, made organising the event an especially challenging ordeal for the department and Shawn.

Within the timespan of 2 weeks, the team went from ideation to submitting the event proposal and co-ordinating with lecturers. The event was slated to have a maximum registration of 100 people, presenting another challenge that is the lack of manpower to manage said participants, since we are just an SIG.

"Hey, do you wanna help out for this one thing?"

Those kind friends of the NYP AI committee who answered yes to the above joined a makeshift event committee to help us out and mainly be Group Leaders (for each group of participants) at the event. Thanks to an intricate lineup of games, the committee was all hands on-deck to source and prepare logistics in order to ensure that the event ran smooth.

Most event organisers count their lucky stars to have half the actual RSVP during the event as registered on paper. We thought we would be most event organisers, turns out we weren't...

D-Day

The committee rolled in an hour early before event start on a rather warm afternoon. Luckily, we had the respite of the well air-conditioned P1 Convention Centre.

After fixing up slides, checking AV, and spreading out logistics, Shawn ran a dry run for the group leaders to ensure everyone was on the same page. With moments to go before doors opened, I (Prakhar), being the emcee, picked up the mic, blasted upbeat party tunes on the reverberating speakers of the hall and welcomed participants with a smile.

Eyes glued to the group layout on screen, an avalanche of students streamed in towards their respective group leaders and took a seat. To our pleasant surprise, 90 people were in attendance, ready for an action-packed afternoon.

Icebreakers

First up on the menu: some quintessential icebreakers. Groups were left on their own to play a few icebreakers like Name Splat, Enchanted Smackdown and Broken Telephone. Thanks to the exuberant energy of our Group Leaders, the ice melted like it was at Sahara and we were pleased to see lots of laughter and joy.

Some groups even combined to play icebreakers together and enjoyed even more ridiculous moments (one team went from flying to Zumba??).

Delectable snacks like chocolate eclairs were served afterwards for everyone. But, the best part is yet to be, of course.

AI Fundamentals

I introduced some fundamentals of AI to students. Although very high level, these elementary concepts would help de-mystify the AI buzz and give students the much needed tangible understanding of what AI actually is. And that's math.

Beginning with the age-old two types of Machine Learning – Supervised and Unsupervised Learning – I explained the purpose of different kinds of algorithms like decision trees and clustering before giving examples of where they can be used in real life as well. Moving on to the topic everyone's really excited about, I laid out the 101 on Generative AI and the popular tools available on the market people can use.

I encouraged participants to take special note of the brief content as they may need to apply it sooner than they think.

Mass Games

AI Sorcerer Showdown

We wanted to make participants engage with each other instead of staying in their silo group. So, we introduced AI Sorcerer Showdown.

In this mass game, participants were given sashes (strips of garbage bag – we were on budget) that they tug on to their pocket. When the 10-minute timer starts, they can start running around the hall to snatch the sashes of other players to secure a "kill" and thus a point for their group.

Upon being killed, players go to their Group Leader who will quiz them on something that was covered earlier in the AI Fundamentals section. If they answer correctly, they get a second life to get back in the game and try again! The group with the most number of sashes at the end wins!

After a briefing and some strategy time, we started the clock and watched some incredibly crafty footwork happen in the hall. Some teams stayed close in solidarity while others went out to get kills. Killed players that payed attention got back in the game and some stayed till the end as well. Once again, the room lit up with joy and we were pleased to see everyone having fun!

Doodle Dash

Then, everyone settled down and we moved on to the next mass game: Doodle Dash.

In their groups, one phone was passed around where each person had to as quickly and accurately as possible draw something they are told to draw so that an AI could recognise it. Sound familiar? That's right, the website is https://quickdraw.withgoogle.com

Group Leaders counted the number of drawings that got recognised within 5 minutes, adding to their group's points secured in the AI Sorcerer Showdown earlier.

Conclusion

After collating the group points, we announced Cheng Hock's (one of the group leaders) team as the winners! We took a quick photo with the team and their prize, a box of shiny Ferrero Rocher. After a series of QR codes on screen promoting NYP AI, we invited participants to give us feedback on the event.

New Leadership

Allow me to introduce our new EXCO and structure of AY24/25 with this opportune blog post:

• President: Wong Wen Bing (DAAA)
• Vice-President: Prakhar Trivedi (DIT)
• Club Secretary: Cheng Hock (DAAA)
• Head of Events Department: Sarah Zoe Sung (DAAA)
• Head of Community Engagement Department: Shawn Goh (DAAA)
• Head of Publicity Department: Faith Yeo (DAAA)

A huge thank you to the previous EXCO for your contributions to the club and for trusting us to pave the way for more exciting things we can do. Thank you to all the volunteers who helped us out in organising the event; your contributions are much appreciated.

The event was a huge success by any measure: we welcomed a lot of new members and promoted the club among freshmen. NYP AI is glad to have pioneered this first-of-a-kind event among SIGs and we are excited to see what other SIGs might organise with this event type. We hope the event encourages SIGs to step out of the traditional bounds of SIG activities like workshops and more to benefit the student body even more.

Get ready for more from us as well. There's a lot of exciting things we have lined up and we can't wait to see you there.

Photo credits: NYP Photography Club and NYP AI Publicity Department

Join Us | Website | Instagram | Discord | Telegram

Tech week is back this year and we have come up with a better and more interesting topic that has been taking the world by storm - “Generative AI”. The workshop was help on day 2 of tech week and physically at NYP after the covid-19 break.

The aim of the event was to educate participants on the basics of generative AI as well as introduce them to the many open source models that they can use in their projects and how they can quickly finetune them to answer questions based on their desired dataset.

The event began with us giving a very quick introduction to AI, machine learning and deep learning and where generative AI fits into all this.

After this we gave a brief intro to generative AI, which is the branch of AI that can create generate new content from the data that it is being trained on. The most popular example of this is chatGPT which has been trained on a large corpus of text and is very good at text generation. Similarly we have models that can generate images (https://huggingface.co/spaces/stabilityai/stable-diffusion), music (https://huggingface.co/spaces/facebook/MusicGen) and now even video (https://huggingface.co/PAIR/text2video-zero-controlnet-canny-arcane). Feel free to play around with these if you like, they are all open source and freely available to the public.

we also spoke about 3 different generative models which were Generative Adverserial Networks (GANs), Denoising Diffusion Models, and Variational Autoencoders.

Some of the key points that were highlighted can be seen in the table below:

TEXT GENERATION

After talking about how generative AI models work, we moved onto talking about text generation which was a part of Natural Language Processing (NLP). AI uses NLP to make sense of textual user inputs by analysing the grammatical structure and of sentences and individual meaning of words, then uses algorithms to extract meaning and deliver outputs.

Coming back to text gen, we spoke about how the next word to be predicted in the sentence will have a high probability of occuring after the word by analysing the patterns and relationships within the training data.

However, this approach doesnt work especially if we want to preserve the context of the sentence. The problem was that models couldnt remember long sequences of text and this was due to something called fixed length context vectors. This is where we introduced the transformers concept.

The transformer has something known as the attention mechanism which allows the model to focus on specific parts of the input and give more weight to them when producing an output. Multihead attention was where there would be multiple focus points at once which captures different aspects of information. Finally positional encoding would ensure that the model can recognize the word order.

The transformers mode was introduced in a paper titles “Attention is all you need”2017 and has come a long way since then with many models such as the GPT and BERT models being transformer neural networks.

Improving Text Generation Models

To improve text generation models there were 3 ways we could have done it.

• The easiest would have been prompt engineering, finding a suitable way to prompt the model so that the desired output can be achieved check out prompt engineering guide (https://www.promptingguide.ai/techniques/zeroshot) which is a very guide resource talking about concepts like zero shot, few shot and chain of thought prompting.
• The second would be Fine tuning which is the process of adapting pre-trained models to perform specific tasks or operate within particular domains. Techniques like quantized low rank adapters (QLoRA) and Performance Efficient Fine Tuning (PEFT) can be used. However this is a very time consuming and computationally expensive process
• The third is retrieval augmented generation which uses both retrieval-based and generative models to produce more informed and coherent responses. It can be explained in 4 simple steps:

1. Query Formulation:

• The process begins with the formulation of a query based on the input received. This query is designed to retrieve relevant information from a given dataset or knowledge base.

2. Information Retrieval:

• The formulated query is then used to fetch relevant documents or data snippets from the external dataset or knowledge base. This is typically done using a pre-trained retriever model which ranks the documents based on relevance to the query.

3. Contextual Merging:

• The retrieved information is combined with the original input to form an augmented context. This augmented context now contains both the original input and the additional information retrieved from the external source.

4. Response Generation:

• A generative model, such as a Transformer-based model, is then employed to generate a response based on the augmented context. This response is expected to be more informative and accurate as it's based on additional external information.

RAG practical

This practical was designed to shows users how to use retrieval augmented generation on their own dataset. For the practical we used the arxiv research papers dataset. The first step was to convert the text to embeddings. To do that we needed to craete an embedding pipeline. We used sentence transformers model to do this for us.

And then we had to use this pipeline. We can see that the embeddings have been produced for the 2 sentences in the docs list.

To store the embeddings we had to use a vector database. For this we made use of pinecone which had a very generous free tier (https://www.pinecone.io/).

The vector index, a data structure used to organize and search vector data efficiently, was created.

Our dataset was of ArXiv research papers. Below is the code to load the datasets.

This data was then processed to have 3 key components: Ids, Embeds, and metadata. This was then inserted into the vector database.

We then made use of langchain for the retrieval QA generation.

After all that code it was finally time to run it to see if it worked. Below is the comparison of the same response with and without RAG.

Image Generation and Captioning Practical

After text generation we moved on to our image generation practical. The goal was to build a simple interface using gradio where users could get the caption of the images by uploading them or type in a prompt to generate images.

Text to image model: https://huggingface.co/runwayml/stable-diffusion-v1-5

Image to text model: https://huggingface.co/Salesforce/blip-image-captioning-base

We made use of the model endpoints and made a simple interface that looks the one below:

Gradio interface code for text captioning:

Gradio interface code for image generation

As the event wrapped up, many participants expressed how engaging and informative they found it. Even with a smaller crowd due to the in-person format, the event allowed for easy clarification of doubts and valuable networking among attendees. The hands-on practical sessions were a highlight, as everyone could follow along, making the learning experience both enriching and interactive.

Links:

Workshop Materials: https://drive.google.com/drive/folders/1MHdWMrzNQ_au4drMEjYbH1Ukgcs-prsr?usp=drive_link

NYPAI Instagram: https://www.instagram.com/nyp_ai/

NYPAI Linkedin: https://www.linkedin.com/company/nyp-ai/mycompany/

NYPAI Membership Form: https://forms.gle/H3C4afosUJVvBgKC6

In July, NYP AI made waves with our "Fun with AI 2023" workshop. What set it apart? Not only was it one of our largest gathering with 80+ participants, but it was also tailored for beginners. Even those without a coding background found themselves diving deep into the world of AI, discovering the technology that's reshaping our world.

Games, AI Ethics, and More!

Before diving into the hands-on session, the workshop set the mood right with some intriguing games. One such game required participants to distinguish between AI-generated and real images. It was not just fun but served a dual purpose to show the astonishing capabilities of AI in mimicking reality and how it should be used responsibly.

This game sparked insightful discussions about AI’s potential, its limitations, and the responsibilities of those harnessing its power.

Concepts Covered

Before we moved on to the hands on portion of the workshop, we covered some of the theory on some basic AI and computer vision concepts and how everything is related together.

Since we were more focused on the computer vision part, we spoke more about some of the tasks that can be acomplished such as Image Classification, Object Detection, Optical Character Recognition (OCR), etc.

In any machine learning project, we start with data collection. This data can be messy, so we clean and organize it through preprocessing. For computer vision, preprocessing means adjusting image sizes, fixing brightness, and sometimes flipping or rotating images to get more variety.

Once the data is ready, we train our model. We use the cleaned-up images to teach the model how to recognise patterns and objects. The quality of the data we collect at the start plays a big role in how well our model will work.

While our discussions often revolve around the intricacies of building models, this workshop placed a heightened emphasis on evaluation. After all, it's the quality of evaluation that truly determines the merit of a model, rather than the complexity of the modeling techniques used. The primary focus was on accuracy, but we also introduced participants to metrics like the ROC curve, true positives, and false positives. However, we kept our exploration of these metrics surface-level to avoid overwhelming the participants.

Following our discussion, we challenged participants to apply their newfound knowledge with a hands-on session using Teachable Machine.

No-Code AI with Google's Teachable Machine

Teachable machine is a tool that offers a no-code solution for machine learning, making AI accessible and understandable to all, regardless of their programming experience. For many participants, it was their first brush with AI, and the simplicity of Teachable Machine ensured they weren’t intimidated. With a user-friendly interface, participants easily created models, teaching their computers to recognise patterns, objects, and more.

While individual experimentation with the Teachable Machine was insightful, the event took a collaborative turn by introducing a group work component. This allowed participants to brainstorm, share ideas, and problem-solve collectively. It was not just about learning the tool, but about collaborative innovation, simulating real-world scenarios where teamwork often leads to. Some of the projects that the participants have worked on were on emotions identifier, among us classifier, dangerous tools detector, and more. After participants finished they then presented their models which allowed others to provide feedback and gain a deeper understanding of the topic.

More of the models created by participants can be found at this link: https://padlet.com/prakhar0706/nyp-fun-with-ai-workshop-submissions-2b6xk1i37i53nl1e

All in all, the feedback that we received from this event was very positive largely due to the hands on portion of the workshop where participants were able to apply their own knowledge on use cases they are interested in. This shows that we should also consider including this for a more engaging event in the future.

During the September break, NYP AI hosted Tech Week in collaboration with the 5 other interest groups at NYP. The goal of this event was to teach Year 1 students the basics of Artificial Intelligence and how it can help people in their everyday lives

The outcome of this event was to give the students clarity on what AI, Machine Learning, and Deep Learning are, and then teach them how to create their own text summarizer using the streamlit library in Python. This web application can perform 2 simple functions

• it takes a link from a website or text that the user types in
• create a text summary using the huggingface transformers library

Inputs:

Output:

The Introduction

We first talked about artificial intelligence, machine learning, and deep learning, followed by the differences between them. For example, one of the differences is that in machine learning you have to do something called feature extraction, which is the process of labelling features. For example, if you want to identify a cat, you need to know what kind of ears it has, what kind of tail it has, etc. With Deep Learning, you do not have to label these features; instead, the Deep Learning model will identify the features for us. After discussing this, we talked about what neural networks are and the 3 different types of neural networks. We then introduced the participants to NLP, which they were able to understand by creating 2 types of text summarizers (abstractive and extractive). Finally, we used our text summarization in a streamlit webapp.

Natural Language Processing (NLP) project

We then talked about NLP and used a text summarization project to show the participants its possible applications. Before we started building the project, we introduced 2 types of text summaries

• Extractive text summarization: Finds the most important sentences in the text and creates a summary by combining these sentences together
• Abstract text summary:  Creates a text summary in his or her own words by looking at the main ideas in the text

The first step of our project was to collect the text data that we wanted to summarize. We wanted the app user to have the freedom to summarize content from whatever website they wanted. So we used the requests and beautifulsoup libraries to collect the text data

The requests library helped us send a request to the website and retrieve the HTML page of the website, while the beautifulsoup library allowed us to retrieve the text from the HTML page essentially acting as a parser.

After collecting all the paragraph tags, we need to extract the text from the paragraph tags and combine them into a sentence. After that, we simply use some regex to clean up the text. This is how we get our final text.

Now we need to do something called tokenization. This involves breaking down the text into smaller units called tokens. We need to create 2 types of tokenizers: Word Tokenizers and Sentence Tokenizers.

After we have created our tokenizer, we need to create a word frequency table. This table, as the name implies, records the number of occurrences of each word in the table. However, this table does not include punctuation or stop words, which are words that do not add any value to the text.

After creating our word frequency table, we find that certain words are more common than others. To reflect the true meaning of a word, we had to perform what is called document length normalization. See the slide below for more information.

After normalization, we need a sentence scores table that shows the importance of each sentence in our text, which in turn allows us to select the most important sentences in the text. Below is the code to accomplish this in python.

Now that we have our sentence score table, we can pick out 30% of the total sentences with the highest score. To do this, we use the nlargest function from the heapq module. The nlargest function takes 3 arguments: the number of sentences as the first argument, the table of sentence scores as the second argument, and the key of this table of sentence scores.

Abstractive text summary

Abstractive text summary is the process of creating the text based on the main ideas. So you can think of it as the AI writing the text in its own words instead of copying the main sentences. For abstract text summarization, we used the huggingface transformers library.

• First, we need to create a pipeline object and assign it a task, which in our case is the summarizer.
• Then we insert our text into the summarizer, set the maximum length of the summary to 400 words
• The minimum length to 100
• Set do_sample to false. This parameter controls what type of decoding takes place.

After that, we simply return our summary text.

Streamlit webapp

Streamlit is an app framework in Python that allows us to build web apps for data science and machine learning. It is very popular because it is very easy to use and compatible with various libraries like keras, scikit learn, tensorflow, etc. Streamlit webapps are written in Markdown, which allows us to easily create good looking webapps. Since we wanted the website to give the user a choice between entering a website link or text, we created a navigation bar with these options.

After that, we simply put a heading and an input field on the page. To generate the text summary, we used the huggingface transformers library to accomplish our task.

By the end of the event, participants were now more familiar with the basic concepts of AI and knew how to use the various tasks in the huggingface transformers library to accomplish various NLP tasks, and they also knew how to use streamilt to create nice-looking web apps.

Afterthoughts

We would like to thank all participants for attending. We hope you were able to learn more about AI and its possibilities. We will be back with more content in the future, so stay tuned!

In June 2020, NYP AI was born. Our vision - To spread awareness of AI among polytechnic students.

Progress (The ups...)

Fast forward 2 years: We've grown to a 100+ member interest group, hosted over 10 school-wide events and ultimately found our place as SIT student's go to for Artificial Intelligence events.

We have our annual flagship NYP AI Summer Camp event, a week long programme introducing students to the up & coming AI concepts such as NLP Transformers and Stock prediction with LSTMs.

More recently, we've delved more into out-of-the-box algorithms, touching upon Tensorflow Magenta, OpenAI's CLIP model, and Huggingface's Transformers, just to name a few.

As champions of open sourcing, all our content (slides and codes) are freely available online. Our Github repository contains all our event source codes, and the content slides for each of our events can be found in the respective readme.md files. Below are the URLs of our Slides (In a Google Drive folder) and our Github Repo.

NYP AI Shared Content – Google Drive

GitHub - NYP-AI/Learning-Materials: A Public Repository containing all of NYP AI’s past event materials

A Public Repository containing all of NYP AI’s past event materials - GitHub - NYP-AI/Learning-Materials: A Public Repository containing all of NYP AI’s past event materials

GitHubNYP-AI

Our blog is constantly updated to feature our latest initiatives & events. For students who are unsure whether Artificial Intelligence is for them, they can read our blog posts to learn more about our past events. Our blog gives prospective members a glimpse into all our events - such as the concepts we've covered and the way we teach these concepts. We even have a "Projects" page on our website, dedicated to showcasing some of our projects from past events - which will hopefully pique the interests of curious students.

And the downs...

But it didn't always go so smooth. At the beginning, we were struggling to make a place for ourselves. Our overemphasis on coding proved a double-edged sword as some students (understandably) found coding AI algorithms too daunting. No doubt Machine Learning requires quite a bit of coding, be it during data preprocessing or model building. However, if done right, breaking into this subject can be an enjoyable and smooth one.

Learning how to teach

We spent the next 2 years "perfecting" the craft of teaching AI, and although we've yet to reach the peak, what we can say for sure is that we're better than our version 2 years ago.

Here are 3 "mindset shifts" we've come to realize, powerful ones which we sure can benefit other interest groups.

Our first mindset shift: Upholding the relevancy of our events.

Our first few events always ended up with a trained model in a python notebook. No websites, no applications - just a .ipynb notebook.

Over time, we realized there were better ways to do this. By creating an application that integrates these AI models, we would accurately mirror real-life deployments of AI. You'll probably never interact with production AI in a python notebook. Instead, you'll interact with AI on your phone (Siri, Facial Recognition) or on websites (Recommender systems, Chatbots) instead. Cutting a long story short - We'll be teaching students how to apply AI in a real-world setting.

Furthermore, by showing students how to integrate their trained models into their own applications, we teach them how to use AI in their own personal/school projects. During our past events, we taught students how to save their AI models onto pickle files and deploy them onto their Flask websites. Additionally, we covered how to "AI Supercharge" their applications using AI REST APIs - granting them access to the latest & greatest models, trained by the biggest research labs. After all, why reinvent the wheel when there are already pre-trained models freely available online?

Finally, by having an application as a deliverable, it serves as a cool end-product, a representation of them having successfully applied AI. It's this sense of accomplishment which we hope will spur students to delve deeper into AI.

Our second mindset shift: Concepts over code.

Coding proved to be a high barrier to entry for students looking to break into AI. Our initial efforts of wanting to provide an in-depth view of AI backfired when we realized that some students could not catch up with the coding exercises.

We decided to shift away from our "build from scratch" mentality which led to a lot of (unnecessary) code writing. Instead, we shifted to using interesting Github packages and projects (Huggingface models, CLIP model, Tensorflow Magenta, etc..), drastically reducing the amount of code written while maintaining the integrity of our events. Using these packages, we could easily illustrate AI concepts with highly abstracted pieces of code, ensuring that our attendees wouldn't get "burnt out" from having to type out long lines of code.

More recently, we've come to embrace no-code solutions. Highly-visual tools such as "Teachable Machine" were excellent in teaching students how to train their first Image classification model. The simple drag-and-drop interface made learning a very painless and interesting experience for first-timers. Other forms of no-code solutions include drag-and-drop model builders such as "Microsoft Cognitive Toolkit" which allow us to teach data preprocessing and hyperparameter tuning easily.

To put it simply, we came to realize that coding was merely a means to an end; A way for students to gain a better understanding and appreciation of AI.

Our third mindset shift: Publicity 👀

An event is only as good as the number of participants it benefits.

Despite hosting a variety of events, we realized that NYP AI still wasn't very well-known within NYP SIT. Despite many students taking an interest in AI, a substantial number of them had never attended our events.

We realized that more had to be done on our end to publicize NYP AI and our events to students. Knowing this, we utilized social media (Instagram), enlisted the help of our teachers and academic club (SIT Club), and used word-of-mouth to increase our mindshare. We organized inter-SIG events such as "Tech in SIT", a collaboration between all SIT SIGs to host a week-long worth of events. We organized school-wide puzzles during "Solve IT", a puzzle campaign aimed at teaching students new concepts through gamified puzzles. With the support of our school & teachers, we used NYP open houses & orientations to publicize NYP AI (SIGs in general). Moreover, we decided to introduce AI concepts in a highly beginner-friendly way (As explained in our second mindset shift), which greatly expanded our target audience.

Our efforts were not in vain. With a more "offensive" publicity strategy, we've garnered more average participants for our events. We've managed to reach out to other NYP schools such as NYP School of Engineering. Given the interdisciplinary nature of AI, we're looking to expand NYP AI to other NYP schools in the future, bringing room for more all-rounded collaborations.

Epilogue

We've made it far, but not without the passion & commitment of the NYP AI leadership team, comprised of six - Jun Cheng, Tony, Dylan, Jing Kai, Nuzul and me. We're proud of what NYP AI has become and we're excited for what NYP AI will be as we hand over the torch to our next leadership team.

~ Sincerely
The NYP AI Team (20-22)

Let's Talk NLP

GitHub - NYP-AI/Learning-Materials: A Public Repository containing all of NYP AI’s past event materials

A Public Repository containing all of NYP AI’s past event materials - GitHub - NYP-AI/Learning-Materials: A Public Repository containing all of NYP AI’s past event materials

GitHubNYP-AI

On 16th July 2022, NYP AI hosted its 2nd NLP event where participants had a great time creating their own Spam Classification Model. We covered an intro to Machine Learning, Data Preprocessing, Training our own NLP model, Deploying the Model with Flask and using Hugging Face APIs for semantic Analysis.

We kickstarted the session with a quick introduction of Machine Learning and we covered the different types of Machine Learning- Supervised Learning, Unsupervised Learning and Reinforcement Learning while highlighting examples for each of them and the differences between them allowing members to get a good grasp on Machine Learning as a whole.

After the introduction of Machine Learning, we went through the Spam Dataset and explained the columns. The Dataset had a total of 6k Data and they will be fed into the model.

After getting a clear understanding of what the spam.csv contained. We proceeded with Data Preprocessing where we used pandas to remove the remove extra columns and rename them. After that, we used the Countvectorizer from Sklearn to convert words into vectors.

Once we preprocessed the data, we split the data into a train and test set. From there, we proceeded to train our model using Logistic Regression from Sklearn as well. Then we tested out our model using the data from the test set.

Lastly in the pipeline overview, we moved to Generating Predictions. As a human, we tend to make mistakes and it would take us a long time when we try to analyze a huge data set and make predictions, however by using ML we can save a lot of time and there will be close to 0% errors. The participants also got the chance to create their own text sentence and check out if their sentence is Spam or Ham by using the model and vectorizer from earlier.

It would be very inefficient if we had to train the model every time we wanted make a prediction. To eliminate this and provide persistence, we utilized Pickle to save our CountVectorizer & Logistic Regression instances.

Using Flask to deploy our models

To invoke the spam text classifier using our flask website we will create an app route and a function which does 3 things.

1. Collect the text entered by the user in the input field
2. Convert this text into a vector using our vectorizer pickle file
3. Predict whether this vector is spam or not
4. Return the result to the website

The code below does these steps for us

Using Huggingface API endpoints 🤗

We used the hugging face sentiment analysis API to classify whether the given text is positive, negative or neutral.

https://huggingface.co/finiteautomata/bertweet-base-sentiment-analysis

This is the API we will be using.

To use this API we need to click on the link go to deploy and then click on accelerated inference. After that, we copy the pre-generated python code and paste it into our python file. If you are new to hugging face you will also have to create an API endpoint, to do that you may follow the steps given in the link below. (https://huggingface.co/docs/api-inference/quicktour)

After doing that we will print the variable output to check if we imported our API correctly.

[[{'label': 'NEG', 'score': 0.0017993206856772304}, {'label': 'NEU', 'score': 0.0065250033512711525}, {'label': 'POS', 'score': 0.991675615310669}]]

So with this in mind, we will pass in our own sentences and have it do the same thing. So we will be invoking this API using our flask website. To do this we will be creating a function called sentiment() that will be accessible using the /sentiment route.

This is our output that is rendered on the website

However, this output might be a bit hard to understand for the user so we will be outputting it to make it more understandable for the user.

To do that we will create 2 empty variables pos_output and neg_output and assign the positive and negative scores to these variables respectively. The output will also be in percentages.

The output now looks something like this

Challenges faced while working on this project

One of the challenges was that the flask app would not work properly when using virtual environments and we had to manually keep killing the processes on the flask port. To solve this problem we used conda env and also set the app to run on a specific port.

GitHub - NYP-AI/Learning-Materials: A Public Repository containing all of NYP AI’s past event materials

A Public Repository containing all of NYP AI’s past event materials - GitHub - NYP-AI/Learning-Materials: A Public Repository containing all of NYP AI’s past event materials

GitHubNYP-AI

Intro To AI - This weekend on 14th May 2022, NYP AI hosted its second event covering the basics of AI to spark interest in beginners. We covered where Machine Learning can be applied, the 3 main areas of Machine Learning, Training a Image Classification Model with Teachable Machine and more...

After learning about real-life examples of machine learning through discussions with the help of Mentimeter, we introduced Machine Learning in a way that's easier to comprehend by comparing it to Traditional Programming.

Then, we covered what features and labels meant in the context of machine learning with the example of Predicting Housing Prices. Once again we went back to Mentimeter for further discussions on examples of features and labels, where Email Body is the feature and whether it's spam or not is the label in the context of Email Spam Classification.

Additionally, we went through the 3 main types of Machine Learning, Supervised Learning, Unsupervised Learning and Reinforcement Learning highlighting examples for each of them and the differences between them allowing members to get a good grasp on Machine Learning as a whole.

Finally, we ended the first part of the session with the Machine Learning Pipeline to encapsulate what was learnt and show how it applied efficiently in the real world. At this stage, participants were able to ask questions to clarify their doubts about any part of Machine Learning.

Computer Vision

For Computer Vision, we started with what's Computer Vison about and explained Image Captioning and Optical Character Recognition. We also gave an example of how the image in the slide would be read if we used Computer Vision. From here we moved on to how an image is represented in a NumPy array to computers.

For Grayscale Images; 0 corresponds to black, 255 corresponds to white and anything in the middle is grey.

For RGB images, each pixel has its own NumPy array which contains [Red Intensity, Green Intensity, Blue Intensity] which also contains the range of 0 - 255. At this stage, members were able to ask questions to clarify their doubts on any part of Computer Vision. Even though the introduction to Computer Vision was short, it was succinct participants were able to learn the essence of it.

Image Classification

So, what is image classification? Simply put, it is a supervised learning technique to train a model to recognize labelled pictures.

The following pipeline below illustrates the steps in image classification:

To start off, the first step would be to gather data. If we want to train a model to classify between cats and dogs, you would have to download some images of those pets from the internet.

Afterwards, we also need to define the labels for each class:

Second, we will make use of the data we gathered earlier to train the model.

A model is trained by extracting key features of an image.

Suppose that we are training a model to classify a shoe, when the picture is processed, it extracts features such as the edges of the shoe. By doing so, the model will use that to determine whether an image passed on to it, belongs to a shoe class or not.

Now that the model is trained, we are ready to test the model! Let us take the earlier example of classifying cats and dogs. To test the model, simply upload new images of these pets and see if the model can classify those images correctly.

Teachable Machines

It's time to put the knowledge we acquire into practical use. We are all familiar with the childhood game "Scissors, Paper, Stone", in this session, we will use teachable machines to train an image classification model to differentiate between paper and stone.

Teachable machine is a platform which makes it easy to create your ML model without any coding any line of codes, all you have to do is to click a few buttons and you are ready to go with your very first model :D

Click on the following URL to the image classification page:

https://teachablemachine.withgoogle.com/train/image

The interface can be broken down into 3 sections, starting from the left, it is where you define your classes and upload your data into the model.

Click on the pencil icon to edit the name of the class, change it "Rock". Similarly, change the name of the second class to "Scissors".

The next step would be to turn on your webcam, and click on the "webcam" button. By turning on our webcams and forming the shapes of paper and stone using our hands, we can input these images into the model in real-time.

Afterwards, to form the shape of a rock, raise up your hand and clenched your fist. Click on the "hold to record" button to begin collecting data.

It is recommended to capture a similar number of data, if you capture 100 sets of rock data, try to capture around 100 sets of paper data as well.

*Note to rotate your fist at different angles so as to capture a more diversified set of data. To diversify the data, even more, you can move it around your hand instead of statically leaving it in the same position*

After collecting data for both rock and scissors classes, the next step would be to train your model.

Under the “Training” Tab, click on the “Train Model” button to start training your model.

Important: You must leave the tab open for the model to train

For the final step of this practical session, we have to test the model to determine its accuracy in classifying the different classes. Form the shape of stone and paper and rotate them at different angles or move your hand around in a different position to test it.

The output section is the confidence level of the model in classifying the classes.

If the model is not accurate in classifying some variations of the data, do consider adding more diversified data to improve its accuracy. However, the drawback of adding more data would be that it would take a longer time to train the model.

Hope you have learned something and had fun training your very first image classification model!

We will be back with more interesting content in the future, stay tuned!

GitHub - NYP-AI/Learning-Materials: A Public Repository containing all of NYP AI’s past event materials

A Public Repository containing all of NYP AI’s past event materials - GitHub - NYP-AI/Learning-Materials: A Public Repository containing all of NYP AI’s past event materials

GitHubNYP-AI

Let's Generate

Generative AI - This holiday from the 21st-22nd March 2022, NYP AI hosted its first event solely focused on AI for generation.  Using the latest, state-of-the-art deep learning architectures & models, we generated images & music melodies using Artificial Intelligence.

Day 1: Fun with Images

Using AnimeGANv2, we are able to generate an "anime" version from any given image. AnimeGANv2 is a Deep Learning model combining elements of Generative Adversarial Networks & Neutral Style Transfer.

Using pre-trained models from PyTorch, we're able to implement our first GAN under 7 lines of code...

Diving deeper: We explored CLIP, where Natural Language Processing & Computer Vision converges. CLIP is a SOTA deep learning model boasting exceptional capabilities such as zero-shot classification, image similarity search & much more...

Using the CLIP library hosted on Github, we created two models
1. Zero-shot Classifier: Give a list of textual descriptions & have the model decide which description best fits the given image
2. Image similarity search: Finding the most similar image among a bunch of images, given an input image...

Day 2: Music Generation

Ever wondered how to leverage music for machine learning? Here, we created 2 use cases for music in ML...

Using Tensorflow Magenta, a popular Machine Learning library for art, we used...
1. MelodyRNN to continue a given musical sequence by generating subsequent notes
2. Music VAE to interpolate two musical sequences together (Think of it as smoothly transitioning between song A to song B)

Machine Learning is about the data... Hence, we dived into "Note Sequence" at the start. A Note Sequence is a datatype we use for music, which contains a sequence of notes that form a musical melody.

MelodyRNN is an LSTM (Long-Short term memory) model which is used for continuing a given sequence. Apart from generating music, it's also used to generate text, etc...

Music VAE, short for Variational Auto Encoder, allows us to generate a "transition phase" from music A to B.

By simply specifying how long we want our transition to be, we can generate a new piece, combining multiple songs into one...

That's a wrap

Generative AI is one of the coolest, up & coming applications of AI. We hope that you enjoyed learning about it, just as much as we enjoyed creating this content 😉

Until our next event, stay safe!

~ The NYP AI Team

NYP AI: NLP & Deployment is held in conjunction with Nanyang Polytechnic's Tech In SIT initiative, a collaboration between the 4 special interest groups in NYP SIT.  On the 22nd of December, NYP AI hosted our Artificial Intelligence event: NLP & Deployment. This is aimed at Year 1 students who want to discover their interests in the Tech landscape.

GitHub - NYP-AI/Learning-Materials: A Public Repository containing all of NYP AI’s past event materials

A Public Repository containing all of NYP AI’s past event materials - GitHub - NYP-AI/Learning-Materials: A Public Repository containing all of NYP AI’s past event materials

GitHubNYP-AI

Our end product was simple: A Flask application that performed two functions
1. Clickbait Classification using our own trained model
2. Sentiment Analysis using HuggingFace APIs

Training our Clickbait Classifier

To start off, we covered Text Preprocessing techniques, notably the Count Vectorizer. This was followed by an introduction to the Logistic Regression classification algorithm.

Imagine the inefficiency if we have to train our model every time we want to make a prediction... To provide persistence, we utilized Pickle to save our CountVectorizer & Logistic Regression instances.

Moving on to Deployment, we utilized the Flask web framework. We utilized our trained model to make backend predictions.

Using API Endpoints

Using HuggingFace's API Hub, we were able to tap into state-of-the-art Models, without having to train them on our own.

Using the Requests library, we are able to interface with the API in python

One way for AI to provide value is through exposure in Web Applications. After obtaining our two AI tools - Our trained model & API, we implemented both of them onto a website via Flask, hence allowing anyone to utilize these built solutions.

Afterthoughts

Thanks to all participants who made it - we hope you've learnt something new :)
A shoutout to the 3 other Interest Groups involved in Tech Week 2021
NYP Infosec: https://nypinfosec.blogspot.com/p/about-us.html
NYP DSC: https://discord.gg/jWqkPB9PBn
NYP LIT: https://nyp-lit.github.io/

Until then!

~ NYP AI

NYP AI Summer Camp 2021 is the second AI Summer Camp hosted by NYP AI, held from 20th September - 24th September, 2pm - 4:30pm daily. This time, we dabbled into applicable and interesting AI, including AI Stock Prediction, Face Detection & Tweets impersonation...

GitHub - NYP-AI/Learning-Materials: A Public Repository containing all of NYP AI’s past event materials

A Public Repository containing all of NYP AI’s past event materials - GitHub - NYP-AI/Learning-Materials: A Public Repository containing all of NYP AI’s past event materials

GitHubNYP-AI

Day 1: Computer Vision

Diving into Day 1, we covered the Open-CV library, including how to read images and capture live frames from our computer's camera...

We went on to Object Detection, utilizing CV2's inbuilt CascadeClassifier. From converting images to grayscale, to outputting the bounding boxes, the entire pipeline was coded out.

Finally, we utilized Streamlit to host our model online. Using Github to store our code, and the Streamlit Sharing app, we were able to place our website on the World Wide Web...

Day 2: Natural Language Processing

Starting off Day 2, two preprocessing techniques for text was covered: Tokenization & Word Stemming

We then went on to explore the capabilities of NLP models, using the Huggingface library for NLP tasks including Text Summarization, Text Generation, Sentiment Analysis, Masked Language Modeling.

Named Entity Recognition using spaCy was also covered

Diving deeper, we moved on to Word Embeddings, which most State-Of-The-Art Machine Learning NLP models utilize.

We utilized Tensorflow's Embedding Projector (https://projector.tensorflow.org/) to visualize embeddings in 3D space.

Pretrained Embedding Layers were obtained from Tensorflow Hub, for us to obtain our very own word embeddings.

Finally, to measure the similarity of two words, we utilized the Cosine Similarity function on the word embeddings.

Day 3: Time Series

Stock Prediction with Deep Learning... We used Yahoo Finance to obtain the past 5 Years of financial Data for AAPL stock. We would then predict the closing price.

Today, we introduced preprocessing techniques for Time Series, most notably the Sliding Window method.

The Sliding Window method would allow us to create datasets for Time Series Prediction

In our case, we had a window size of 60, meaning that we would use the past 60 days worth of closing prices to predict the next closing price.

Moving on to Deep Learning, we covered concepts like Layer Types and Activation Functions.

We then dissected the Tensorflow Deep Learning Pipeline: The 4-step process of Defining, Compiling, Training & Evaluating

Finally, we made predictions on our Validation set and compared them to the actual prices (Funnily enough, this was during the Evergrande crisis.. so you could see a steep descent in global prices...) 🤪

Day 4

Here, we explore the malicious use cases of Artificial Intelligence. Our task for the day: Learn to impersonate someone's tweets.

We utilized Snscrape for our Tweets Scraping. By specifying the username, we could extract a specified number of tweets from that person and write them out to a .txt file.

For Text Generation, we utilized Markov Chains.

For our Malicious Use Case, we would replace all URLs in the tweets with our own malicious URL. We would then fit our Model on these preprocessed tweets.

By fitting our model on these tweets, we are able to capture the linguistic style of that user, hence "impersonating" his/her tweets...

Day 5: Recommender System

Plunging into Recommender System, we covered two forms of Recommender Systems: Simple recommender system & Content-Based recommender system

For our Data Ingestion, we utilized the Kaggle API to download our Dataset: The MovieLens Dataset

For our Content-Based recommender system, we covered Text Preprocessing techniques including Bag-of-Words and TF-IDF. We then briefly touched on Matrix Factorization, before moving on to Cosine Similarity.

Finally, we trained our Content-Based recommender system using Cosine Similarity & used it to provide recommendations.

Afterword

We'd like to thank all our participants for making NYP AI Summer Camp 2021 a huge success! Not forgetting our Planning Team: Lim Jing Kai, Loy Jun Cheng, Tony Yu, Dylan Kok, Nuzul Firdaly & Alex Chien. Without them, the Camp wouldn't have been as exciting as we'd envisioned...

We'll continually come up with more exciting & updated AI content, so do keep a lookout for our future events 👀

NYP AI Summer Camp 2021 is the second installation of our trademark Summer Camp Series. It'll be held from 20th to 24th September, 2pm - 4.30pm daily. Be sure to check it out!

We've prepared a whole lot of exciting activities :)  

Here are some of the activities we have in store for you ~

• Stock Prediction with Deep Learning LSTMs 📈
• Face Detection with CV2 😉
• Natural Language Processing with state-of-the-art Transformer architecture 🌐
• Using AI to impersonate someone's tweets 😈
• Deploy your Machine Learning models on your very own publicly accessible website 🧰
• Create your own Movie Recommender System using popular, relevant techniques 🍿

What we'll be covering...

Day 1: Computer Vision Techniques

After this, you'll be able to create your own Object Detection algorithm for just about anything (Fruits.. People...etc.). You'll also see how your camera can recognize your face, how images are broken down by machines, and how to deploy your personal projects to the browser.

You'll gain insights on Image Processing - how machines are able to break down images into number sequences.  Subsequently, you'll train an Image Classification Algorithm right from your browser and camera.

Additionally, we'll be covering the OpenCV library and using its built-in Object Detection Algorithms for face detection. At the end, you will have built your very own Object Detection algorithm, deployed on the web.

Day 2: Natural Language Processing

Dive into the domain of Natural Language Processing: Linguistic interactions between man and machine. After this, you'll understand how a Chatbot works, visualize how Computers understand languages and meddle with some of the most popular applications of NLP.

We'll touch on the core Language Preprocessing techniques: Stemming & Tokenization.

Utilize state-of-the-art NLP models using the Transformers Library. You'll learn how to do Text Summarization, Text Generation and Named Entity Recognition.

We'll also be covering how Machine Learning algorithms understand words - through Word Embeddings.

Day 3: Time Series

Learn how to leverage Machine Learning for Sequential Data. After this, you'll be able to write code to work on sequential information, from Text generation, Sentiment Analysis, and yes, Stock Price prediction.

Learn about Deep Learning pipelines using the Tensorflow library. You'll delve into creating your own Neural Network Architecture, compiling your first Deep Learning Model and training it on your own data.

We'll touch on preprocessing techniques for sequential data, including N-grams and Scaling. By the end of this, you'll have trained your own Stock Price prediction algorithm with the stock of your choosing, with the latest data obtained from Yahoo Finance.

Day 4: Evil AI

AI is a double-edged sword. Learn how AI can be used to commit malicious activities, and how to recognize them. After this, you'll never look at your emails the same way again.

You'll learn how to impersonate someone's tweets: Through generating text with a similar linguistic style using Markov Chains.

Day 5: Recommender Systems

Delve into the technologies used by tech companies. Ever wondered how Youtube videos uncannily pop up in your feed? Interested in finding out how Movie recommendations work? After this, you'll know about the different types of recommender systems and implement one on your own.

You'll learn how to create a content-based recommender system using Cosine Similarity and Matrix Factorization. You'll also meddle with Text Preprocessing techniques, primarily TF-IDF.

I'm sold, now what...

Come sign up for our event: https://forms.gle/xPMMhAanxnzha7jw6

See you there ~

GitHub - NYP-AI/Learning-Materials: A Public Repository containing all of NYP AI’s past event materials

A Public Repository containing all of NYP AI’s past event materials - GitHub - NYP-AI/Learning-Materials: A Public Repository containing all of NYP AI’s past event materials

GitHubNYP-AI

Ever wanted to get into the AI field, but didn't know where to start?

NYP AI has got you covered! Kickstart: AI is an initiative by NYP AI, a 2-day programme teaching introductory Machine Learning code and concepts to participants.  Our event was held across two days, from 21st - 22nd June 2021.

Day 1: ML Concepts & Libraries

To kick things off, we covered key concepts related to supervised learning, terminologies and real-life applications.

This was followed by an introduction to the core datatypes of Machine Learning - Pandas Dataframes and Numpy Arrays.

With a clear overview of Supervised Learning and datatypes, we moved on to the modelling stage. In line with our vision to push for Explainable AI, we went down to the essence of Linear Regression - down to the math. This removes the "Black-box" element of Machine Learning algorithms. Understanding how Machine Learning algorithms work helps in understanding their strengths and limitations.

At the end of Day 1, our participants walked away with fresh insights on the Machine Learning landscape, together with a Housing Price Prediction algorithm to showcase their new skillsets.

Day 2: ML Pipelines

What good does ML do if we can't apply it properly? That's what we served to explain on Day 2. Covering popular approaches to Machine Learning projects, we set forth equipping our participants with the right ML techniques and strategies.

We covered the core Data Pipelines, namely: Data Ingestion, Data Exploration, Data Cleaning, and Feature Enginering.

This was followed by core Modelling Pipelines: Data Splitting, Model Training, Model Evaluation and Hyperparameter Tuning.

With a coherent and comprehensive code walkthrough, our participants were able to consolidate their knowledge.

Till the next Time ~

The team at NYP AI is glad that many participants have gained much knowledge on Machine Learning. We'd like to thank all participants for the overwhelmingly positive feedback and insightful suggestions 🤗. Do keep a lookout on our socials https://www.instagram.com/nyp_ai/ for updates on our latest events :)

GitHub - NYP-AI/Learning-Materials: A Public Repository containing all of NYP AI’s past event materials

A Public Repository containing all of NYP AI’s past event materials - GitHub - NYP-AI/Learning-Materials: A Public Repository containing all of NYP AI’s past event materials

GitHubNYP-AI

NYP AI hosted its first-ever Unsupervised Learning event on 8th May 2021. Armed with detailed slides and concise code, we were ready to introduce a new Machine Learning paradigm to our members, through two of its more popular applications: Clustering and Dimensionality Reduction

To start off, we cleared all misconceptions regarding Supervised vs Unsupervised learning. This involved providing side-by-side comparisons of their definitions, algorithms and applications.

Clustering: K-Means

Moving on to our first Algorithm of the day: Clustering with K-Means. With clear diagrams to aid our members' understanding, we were able to get the point across quickly.

Unsure of how many clusters to initialize? We got that covered too :)

Dimensionality Reduction: PCA

Moving on to a slightly more difficult concept - Dimensionality Reduction with PCA. Notwithstanding the tough visualizations required to understand, we were able to paint a clear picture using simple diagrams.

What better way to understand something than to apply it 🤷.... To end off our theory sessions, we implemented these exact algorithms in code. Having physically implemented these algorithms, our participants were now confident in using them for their own datasets.

Till then ~

Not as hard as it seemed, eh? Hoped everyone managed to learn something from today's session. Until then ~ 😊

GitHub - NYP-AI/Learning-Materials: A Public Repository containing all of NYP AI’s past event materials

A Public Repository containing all of NYP AI’s past event materials - GitHub - NYP-AI/Learning-Materials: A Public Repository containing all of NYP AI’s past event materials

GitHubNYP-AI

NYP AI CV Camp 2020 takes a deeper dive into the Computer Vision landscape. From Image Processing to Deep Learning Convolutional Neural Networks, we cover the niche aspects of Machine Learning. The camp was conducted across four days: 4th - 7th January 2021, including 2 days of learning and 2 days of assignments.

Day 1: Image Processing

To start off the Camp, we reintroduced the already familiar concept of Supervised Learning. In the later stages of our Camp, we would be heavily focused on supervised learning modelling techniques, hence we emphasised the need for a solid understanding of the Supervised Learning paradigm.

This was followed by a high-level overview of the Image Classification domain.

To offer participants a code-free way to train their own image-based models, we utilized Google's Teachable Machine, which you can find here @ https://teachablemachine.withgoogle.com/.

Teachable Machine allows you to train your own Image Classification model, straight from your webcam.

Diving into the details of Image Processing, we covered Numpy Arrays for image datasets.  

Participants also got to create their own image from scratch, using Numpy Arrays.

All these skills culminated in the final project of the day: Training our own Fruit Classifier.

Day 2: Convolutional Neural Networks

For our Second Day, we introduced state-of-the-art Machine Learning Algorithms for Images: CNNs.

By explaining the limitations of Traditional Algorithms for Images (I.e Logistic Regression), participants could see the clear benefits of utilizing CNNs.

Given the Black-boxed nature of Deep Learning Algorithms, concise animations were used to provide a deeper level of understanding of the CNN algorithm. Namely, the concept of Convolutions and Filters were introduced.

As CNNs allow us to work with 3 Dimensional, coloured images, our preprocessing steps would slightly differ from Day 1.

Moving on to the more exciting aspects of Day 2 - Actual modelling using Tensorflow's Sequential API

The results of our Algorithm were extremely accurate - A gigantic leap from yesterday's Logistic Regression Model.

Moving Forward

We really had a blast teaching our members new ML techniques, particularly dabbling with Deep Learning Algorithms. We'll be back with even more awesome content in the future🤞. Until then! ~