We scan new podcasts and send you the top 5 insights daily.
Akash Kulgod noted that artificial intelligence models can process visual and auditory data but cannot smell because no internet of smells exists. Dognosis plans to record dog neural representations of odors to build a perceptual smell dataset, creating reinforcement learning from canine feedback to finally teach AI olfaction.
Akash Kulgod shared how Dognosis partnered with Singapore security forces to deploy canine wearables after explosions. Rather than relying on a handler visually reading a dog, the harness pairs GPS tracking with real-time EEG readouts to automatically project chemical residue heatmaps across a full 3D spatial model.
Akash Kulgod pointed out that analytical chemistry standards like GCMS detect volatile compounds at parts per million, whereas canine olfactory bulbs detect parts per trillion. This sensitivity difference allows dogs to spot stage 1 and 2 cancers easily while ignoring external confounds like recent smoking or specific diets.
Today's AIs are trained on the final product of human cognition (e.g., articles, code). The next great leap could come from training models on the actual neural "traces of thought," potentially via technologies like Neuralink, to solve for creativity and unverifiable domains.
Akash Kulgod explained that training dogs to give binary physical alerts like sitting or pawing takes 12 months and loses nuanced data. Replacing physical actions with brain-computer interfaces cut training down to eight weeks. Sensors capture brainwaves and heart rates to measure confidence, focus, and ambiguity directly.
Applying AI to biology isn't just a big data problem. The training data must be structured for reinforcement learning. This means it must be complete (including negative results) and allow for a feedback loop where AI predictions are tested in the lab, and the results are used to refine the model.
Digitizing smell has been impossible until now because the human nose has over 300 sensory "channels," compared to just three for color (RGB). This complexity required mature AI to create the high-dimensional "map" needed to interpret and organize scent data, a task too complex for previous technologies.
The efficacy of cancer-detecting dogs lies not in identifying a single biomarker but in recognizing a complex, irregular pattern among thousands of emitted chemicals. This suggests that creating an artificial 'nose' for diagnostics requires modeling complex systems, not just searching for a specific molecule, a task well-suited for AI.
To overcome the criticism of individual animal variability, researchers developed a novel framework. Each sample was evaluated by multiple dogs, and a Bayesian model weighted each dog's input based on its historical performance. This created a stable, aggregated score, ensuring standardized and replicable results even if one dog performed poorly.
Akash Kulgod observed that dogs naturally experience off-days roughly 10% of the time, degrading test consistency. Dognosis addresses this by having five dogs sniff each sample while AI models track neurological stress and engagement. The system dynamically downweights scores from unfocused dogs to produce a reliable consensus.
Humans lack the precise vocabulary to describe abstract senses like smell. Google's AI for Estée Lauder overcame this by building a structured framework connecting ingredients to technical categories and then linking them to evocative, emotional descriptions, making the abstract understandable and marketable.