Engineer's View of the Brain: From Noisy Inputs to Internal Activation Perturbation

Version 1 (exploratory preprint). Revisions in progress. Summary When a system produces a different output than the one intended, does the cause lie in the information it receives, or in disturbances inside the system? This paper explores that question with two small experiments on pretrained neural networks. It starts from an everyday slip (“I’d like to park my cat” instead of “car”) and asks what can be measured. What was done Noisy inputs. I added noise and blur to images and tracked how an image model’s internal representation (DINOv2) and a separate classifier’s predictions changed. Blurring both a cat and a car image brought their representations closer together (distance 66.8 to 36.4; cosine similarity 0.02 to 0.69). Neither blurred image came close to the other’s original representation, and the classifier still gave them different labels. Noise inside the model. I kept the input fixed and added noise to a small language model’s internal activations (SmolLM2-135M). Whether the predicted next word changed depended on the layer, the noise size, and the prompt. For “I would like to park my”, layer 29 kept the original prediction (“car”) in 930 of 1,000 trials at a moderate noise level, while layer 28 changed it in 96.8% of trials at a higher level. A one-step disturbance during text generation also redirected the rest of the sentence. What this is notThis is an exploratory study with one image pair, three prompts, and one small model. It makes no claim about human speech errors, stress, or exhaustion, and no brain data were measured. What’s nextMore prompts and models. A test of whether noise-induced errors land on related words (such as truck or van) more often than chance. Repeating the generation experiment at several points. A prediction written down before running each new test. Feedback welcome, especially on better ways to test the related-words question: [email protected] Code: Input-perturbation Notebook · System-perturbation Notebook