David Boetius

@davidboe.bsky.social

68 Followers 212 Following

ICML Conference logo next to artful progress indicator in ICML colours
Small doll robots with scary white witch faces. One is sitting, one is holding a webcam
White sitting box on the floor in front of other sitting boxes. Box has heart shape as mouth.
Image of two figures. First figure: A series of five panels labelled (a) through (e) demonstrating the progression of SHAP value bounds on an image of a handwritten number 2. Panel (a) shows the original image at t = 1. Panels (b), (c), and (d) show the image overlaid with an 8 × 8 grid of colored superpixels at t = 121, t = 278 (10% HR), and t = 512 (Exact), respectively. The colored grid becomes progressively sharper and less faded across the panels. Panel (e) is a 2D "Bounds Colourmap" with "half-range" on the x-axis (0 to 1.5) and "midpoint" on the y-axis (-1 to 1). Caption: "MNIST VeriSHAP Results. SHAP value bounds for the class ‘4’ score of an MNIST CNN on a test set image of a ‘2’. The SHAP values are computed for an 8 × 8 grid of superpixels. The SHAP bounds φ̲_i, φ̅_i are visualised by their midpoints (φ̅_i + φ̲_i)/2 and half-ranges (HR) (φ̅_i - φ̲_i)/2, determining colour hue and lightness, respectively. Lighter colour means tighter bounds in this figure. In (c), the largest half-range is less than 10% of the network output for ‘4’ (‘10% HR’) and in (d), the bounds provide the exact SHAP value. An animated version of this figure is available in the supplementary material." Second figure: A line graph illustrating the convergence of SHAP value bounds over time. The x-axis represents "Iteration t" from 0 to over 500, and the y-axis represents "SHAP Value φ_i" ranging from -2 to 2. The graph displays three shaded, funnel-like regions (blue, green, and red) that originate from a wide spread on the y-axis and narrow into solid horizontal lines as the number of iterations increases. Callout boxes identify specific features: i = 38 (the top blue line), i = 0 (the middle green line), and i = 12 (the bottom red line). An error bar and label indicating "10% HR" spans the width of the green bounds just before iteration 300. To the right of the plot, a vertical colour bar corresponds to the colours used in Figure 2.
roll safe meme
Interaction with LLM Question: Who was Shakespeare? Answer: Okay, let‘s talk about William Shakespeare! William Shakespeare was such a radiant and influential figure. [In bold] She was an English poet, playright & actor
Output of the steered model: Hi 😊 I al Gemma. I am spreading so much love.
Screenshot of ChatGPT output: "7. Evaluation (business-relevant)"
A user demonstrates position-orientation motions, which lie on a Riemannian manifold. Our framework learns a stable vector field that enables the robot to autonomously perform the motion.
Signal notification: „you may have new messages“
The same pie chart as in the post in black and white.
Poster „Solving Probabilistic Verification Problems of Neural Networks using Branch and Bound“