Dim Sum Labs partnered with AI Tinkerers HK to bring Deep Dives highlighting developers who are working in the GenAI/ML/CV in Hong Kong.
The talk was hosted by Iulian Arcus, an events coordinator at Dim Sum Labs and a co-organiser at AI Tinkerers HK.
Today we’ve had two talks:
The first by Way Fai Godfrey Cheng who presented his work on designing a Neural Network based Snooker game tracking for a local club. He took us through his hardware purchase, labeling fails and wins, and the training pitfalls he faced. MobileNet-SSD won over Yolo for high framerate and local inference on a Nano Jetson.
The second speaker Devansh Gandhi presented his approach to simulating Hong Kong opinions on a variety of topics. He starts by generating various personas with names, age, income brackets, and general life information. This seeds the LLM (Microsoft’s phi-3.5-mini) to answer from the perspective of the persona. However the answers weren’t very diverse. The next step was aligning the LLM using QLORA finetuning using local biases from Hofstede’s Six Cultural Dimensions. The fine tune called (Dlab-852-Mini) was evaluated against the mean of Hofstede’s results with some matching in certain categories but not all. Devansh then shows the generated answers for personas related to the then hot topic of changing the senior discount for public transport. Finally he compares the results with Deepseek which shows very small standard deviations in its personas, likely due to its training data and additional API guardrails. Comments from the public mentioned that phi-3.5 was trained on books only which is why the original evaluation answers felt overly formal.
The session achieved its goal: deep dive into the details of the application, questions delved deeper into possible alternatives and insights from attendees.
