Data Basic | Deep Learning projects
🎯 Summary
Podcast Summary: Data Basic | Deep Learning Projects
This 30-minute episode features an in-depth conversation with Professor Sanchesh regarding his career trajectory, the evolution of image processing into deep learning, and the current state and future of computer vision and AI research.
1. Focus Area
The primary focus is the evolution of Image and Video Processing, tracing its shift from traditional signal processing techniques to modern Deep Learning methodologies, particularly within Computer Vision. Secondary topics include the role of Large Language Models (LLMs), Generative AI, data ethics, and cybersecurity forecasting using sentiment analysis.
2. Key Technical Insights
- Evolution of Computer Vision: The speaker transitioned from traditional image/video compression techniques (pre-2012) to deep learning due to the latter’s superior power in solving complex tasks, citing the YOLO (You Only Look Once) model as an example of real-time object detection capabilities unattainable with older methods.
- Current Research Trends: The lab is seeing a rise in the adoption of Diffusion Models for data generation and solving various tasks, alongside the continued power of Convolutional Neural Networks (CNNs) for image/video analysis.
- Data Handling Challenges: A major technical hurdle in video processing is the sheer volume of data required for real-time analysis, which slows down deep learning models. For image processing, limitations often revolve around handling extremely large image sizes.
3. Business/Investment Angle
- Career Opportunities: The field of deep learning, especially computer vision, remains ripe with unsolved challenges, indicating significant future career and research opportunities.
- Industry Transition: The speaker noted that students completing internships in specialized areas (e.g., exercise monitoring apps, medical image startups) often transition directly into employment with those companies, highlighting the strong commercial pipeline from academic research.
- LLM Integration: Future commercial solutions in computer vision are expected to increasingly integrate Large Language Models to leverage vast internet knowledge for tasks like image classification, moving beyond image-only training data.
4. Notable Companies/People
- Professor Sanchesh: The interviewee, whose career spans Mexico, Canada, the US (UC Berkeley postdoc), and the UK (Warwick), is the head of the Signal and Information Processing Lab.
- YOLO (You Only Look Once): Mentioned as a benchmark model demonstrating the power of modern deep learning for real-time object detection.
- ChatGPT: Referenced as an example of how LLMs absorb and propagate existing societal biases present in their training data (the internet).
5. Future Implications
The industry is rapidly moving toward solutions heavily reliant on Large Language Models and Generative AI (including diffusion models). This convergence means future computer vision systems will likely use text-based knowledge from LLMs to augment visual data processing. Furthermore, there is a critical need to address model bias and enhance Explainable AI (XAI) to ensure ethical deployment, especially as data privacy becomes harder to maintain once information is online.
6. Target Audience
This podcast is highly valuable for AI/ML Researchers, Computer Vision Engineers, Data Science Academics, and Technology Professionals interested in the practical evolution of visual data processing and the intersection of technical research with ethical considerations.
Comprehensive Summary Narrative
The episode begins by charting Professor Sanchesh’s extensive global academic journey, which led him from traditional image/video processing (focused on compression during his Master’s and PhD) to the current deep learning paradigm. He notes that the shift accelerated around 2012 when neural networks gained prominence, replacing simpler methods like Support Vector Machines.
The core discussion centers on Deep Learning in Computer Vision. The Professor attributes the rapid adoption of these technologies to their immense power in solving previously intractable problems, exemplified by real-time object detection via models like YOLO. He highlights the visual nature of image data as a key attraction to the field, citing research involving the automated detection and segmentation of specific cells in microscopic bone marrow images, which offers significant time savings for pathologists.
The conversation then pivots to current research in the Signal and Information Processing Lab, detailing projects such as anomaly detection in CCTV footage (e.g., identifying vehicles on pedestrian sidewalks) and predicting future human movement across camera feeds for security applications. A more speculative project involved predicting a person’s appearance 20 years in the future from a single image.
A significant portion of the dialogue addresses challenges and ethics. The primary technical hurdle discussed is the computational load of processing massive video datasets in real-time, though the Professor is optimistic about near-future solutions. Ethically, the discussion tackles model bias, emphasizing that models trained on historical data inherit societal biases. The solution proposed is a strong focus on Explainable AI (XAI) to scrutinize model decisions and ensure fairness, as treating models as black boxes makes ethical auditing impossible.
Finally, the Professor offers career advice, stressing the necessity of a strong foundation in mathematics and statistics (particularly understanding regression as the core of most models) for anyone entering deep learning. Looking ahead, he predicts that LLMs and Generative AI (especially diffusion models) will soon become integral components of mainstream computer vision solutions.
🏢 Companies Mentioned
💬 Key Insights
"For that, I think exploring the ideas from explainable AI is very important. This area allows you to design models that eventually can be analyzed and scrutinized very carefully to justify the decisions."
"What we are noticing now in our lab and our research is that models based on something that is called diffusion are becoming very popular. Diffusion models are those that can also generate new data."
"Now most of the computer vision solutions that we have will also rely on large language models to operate by exploring all the information that is available on the internet to support the process, for example, classifying images."
"Fortunately, when you train a machine learning model with training data, which is our case, even for things that would predict in the future, we have to train the model with actual data that we have from now. Sometimes that data already has biases created by humans, by us. So you have to be careful because the model will tend to learn those biases."
"One of the interesting things we found is that when users tend to be angry, those promoting cyberattacks actually perform the cyberattacks because somehow they feel they have the support of people."
"But one of the main challenges is when dealing with videos because with videos, you have much more data. This is like putting together thousands of images and trying to analyze them. So especially if you want to work with real-time analysis, you see deep learning, that becomes a real challenge. The amount of data you have to process is quite big, and that unfortunately makes things slower."