AI summary by TwinMind
Human Stories in AI: Simon Stochholm
StatQuest with Josh Starmer
37m 27s
Early Programming Introduction and Academic Journey
- Age 10: Father (psychologist) introduced ELIZA software - "ChatGPT of 1980s"
- ELIZA imitated psychologist, asked questions without giving answers
- Fascinated by technology, decided to learn programming
- 1990: No internet available, went to library
- Copied entire programs letter-by-letter from books onto computer
- Some worked, some didn't - had no understanding of what he was doing
- Decade later: Decided to study Linguistics at university
- Forgot about programming passion initially
- Discovered computational linguistics during studies
- Had to switch universities to pursue this field
- Epic university journey: Attended 5 different universities total
- First university shut down program after enrollment
- Original professor was fired for drinking
- Replacement professor was on vacation
- Third contact finally helped transfer to different university
- Completed Master's Degree in computational linguistics
Early Career and Speech Recognition Project
- Business opportunity: Man visited class asking about speech recognition expertise
- Simon volunteered despite only theoretical knowledge
- Challenge: Build switchboard for voice-activated person lookup in 6 months
- Had no coding experience since age 10
- Threat of firing provided motivation
- Successfully completed project using Dragon Speech by Nuance libraries
- Had to adapt system for Danish language
- Manually input phonetic pronunciations for Danish words
- Used rule-based algorithms, not machine learning
- Worked at company for 5 years until 2008 financial crisis
- Got fired when company nearly collapsed
Teaching and Skill Development Phase
- Started teaching people with Asperger's Syndrome how to code
- Discovered students were better coders than himself
- Boss provided multiple coding certifications and courses
- Became proficient programmer through formal training
- Applied to UCL Denmark 9 years ago, been there since
Transition to Machine Learning at UCL
- First 4 years: Taught computer science
- Gained interest in machine learning through Data Camp (free for academia)
- Discovered StatQuest YouTube channel - still uses videos for students
- Found Fast.ai course - "coin really dropped"
- Realized machine learning wasn't impossibly difficult
- Thought you had to be "really brilliant" but discovered average people could do it
- Teaching progression: Now teaches introductory applied deep learning course
- Evolution: Started hands-on, becoming more theoretical and mathematical
- Studying algorithms and math more deeply
- Questioning architectural choices (e.g., U-Net skip connections)
- Found skip connections didn't help on his dataset
- Discovered others had same findings
- Questioned concatenation vs addition in skip connections
Project 1: Ferry Company Energy Optimization
- Client: Ferry company transitioning from diesel to electric engines
- Goal: Optimize harbor time for battery recharging
- Problem: 8 minutes too short for recharging, 17 minutes delays next trip
- Target: 14 minutes optimal harbor time
- Method: Time series to image conversion using Gramian Angular Fields and Markov Transition Fields
- Technique learned from Jeremy Howard in Fast.ai
- Creates colorful images representing time series data
- Used convolution neural networks to identify patterns
- Architecture: Simple autoencoder
- Convolution → Max pooling → Convolution → Max pooling (repeated)
- Downsized images to smallest possible latent representation
- Trained to identify similar images
- Results: Successfully identified ship captains based on thrust patterns over time
- Could predict who was sailing based on thrust usage patterns
- Created shift schedule, compared with actual schedule
- Some captains easily identifiable, others confused with similar operators
- Hardware: Google Colab with single GPU
- Training time: Couple hours for 5 years of data
- Image size: 512x512 pixels, simple images
- Published articles about the technique
Project 2: Medical Imaging for Amputation Prevention
- Background: Denmark had cases of unnecessary leg amputations due to poor blood flow
- Lack of expertise in leg/foot medicine led to amputations instead of treatment
- Became major issue in Denmark
- Goal: Identify blood flow in MRI scans to prevent unnecessary amputations
- Method: Instance segmentation
- Create masks identifying muscles, bones in MRI images
- Use segmentation to locate blood vessels
- Status: Just started couple months ago, waiting for data
- Hardware: "The Beast" - local machine with Nvidia RTX 3090, 10GB RAM
- Plans to move to cloud for full training
- Local machine for debugging and testing
Project 3: Pig Welfare Monitoring
- Origin: Agriculture colleague saw magazine with pigs marked as "angry" or "playful"
- Wanted to implement similar system
- Goal: Monitor pig behavior for welfare assessment
- Track: drinking, eating, lying down, standing up
- Early detection of illness (e.g., diarrhea) prevents farm-wide antibiotic use
- Diarrhea spreads to entire farm, makes meat unsellable
- Setup: Raspberry Pi system with camera recording from ceiling
- Custom case with internet connectivity and large hard drive
- Data collection: Student helper manually annotated all images
- Algorithms tested:
- Started with Detectron2 (Facebook's instance segmentation)
- Switched to YOLO v8 (easier to configure, allows classification and tracking)
- Challenge: Underrepresented classes (drinking, eating, lying down pigs)
- Stable Diffusion experiment for synthetic data generation:
- Dream Booth: Flagged pig images as "nudity" due to pig skin
- Textual inversion: Poor results
- Image-to-image with ControlNet: Limited success
- "Standing pig to sleeping pig" just closed pig's eyes
- No pose models available for pigs lying down
- Control issues: Floor grill pattern caused barbecued pig images
- Generated bizarre images: two-headed pigs, pigs with fins
- Stable Diffusion lacks knowledge of pigs seen from above in farm settings
Project 4: Cloud Movement Prediction for Solar Panels
- Inspiration: Thomas Chapelle (Weights & Biases) article on stable diffusion and cloud movement
- Application: Solar panel company needs 20-30 minute cloud prediction
- Critical for energy production forecasting
- Method: Satellite image analysis for video prediction
- Challenge: Requires supercomputer due to large satellite image data
- High RAM requirements for real-time processing
- Currently only small-scale testing possible
- Goal: Create indefinite time series prediction into future
Advice and Lessons Learned
- Primary advice: "Just do it" (Nike philosophy)
- Key insight: "I'm just an average Joe and I could do it"
- Don't be afraid to try - "What's the worst that can happen?"
- Importance of seizing opportunities when they arise
- Required courage to step out of comfort zone despite being introverted
- Persistence crucial - kept calling until found solution
- Follow your heart and passion
- Be determined and stubborn to reach goals
- Career built on taking advantage of opportunities and persistence
