TinyBrain++: A Compact, Interpretable Alternative to Black‑Box AI
A Mathematical Framework for Structured Data – Not Another Neural Network
Abhishek Thakur
Incoming Student, Computing & AI – The Hong Kong Polytechnic University
TinyBrain++ is a 6 MB, CPU‑only, fully explainable framework for structured data (fraud, churn, anomaly). It uses mathematical expansion (Fourier + Legendre + tensor) and per‑sample gating instead of neural networks. On a fraud detection benchmark it achieves high recall and precision with <0.03 ms inference – no GPU, no black box. This is not a large language model. It is a specialised, efficient alternative for structured data where large models struggle which can become a important step in data processing.
The Hidden Cost of Today’s AI – and What They Don’t Tell You
Large AI models are powerful, but their weaknesses are often hidden behind impressive demos:
Training a single large model can cost $4M–$100M (GPT‑3 class and beyond).
One training run can emit as much CO₂ as five cars over their whole lifetime.
Inference latency is 50–500 ms – too slow for real‑time fraud blocking.
GPU dependency – cannot run on edge devices or standard servers.
Black box – no mathematical reason why a decision was made.
But the most critical gap is rarely discussed: Large models are terrible at structured data.
LLMs tokenise numbers (“8500” → “85” and “00”), losing precision. They cannot reason over rows, columns, or joins. They hallucinate confidently, and one error compounds.
Traditional ML (XGBoost, Random Forest) is mathematically sound but requires thousands of labelled examples, weeks of feature engineering, and months to deploy.
TinyBrain++ fills this exact gap.
Why Neural Networks Struggle with the Same Data
| Aspect | Neural Network | TinyBrain++ |
|---|---|---|
| Number‑handling | Tokenises digits (“8500” → “85”,”00”) – loses value | Preserves full numerical precision |
| Feature interactions | Must learn every combination from data | Explicitly generates pairwise, cubic, quartic terms |
| Training data needed | Thousands to millions of labelled examples | Works with minimal or zero labels (unsupervised start) |
| Explainability | Post‑hoc approximations (SHAP, LIME) | Direct mathematical contribution per feature |
| Hardware | GPU (expensive, power‑hungry) | Standard CPU (ubiquitous) |
MODEL PIPLINE
. Input Preparation Raw structured data is normalised into a consistent format. The system notes available signals – this information later helps calibrate confidence.
Feature Expansion The input is mapped into three complementary mathematical spaces. Each space captures a different type of pattern: periodic behaviour, smooth trends, and cross‑feature relationships. Together they produce a rich representation without requiring large training datasets.
Adaptive Weighting Each input receives a unique set of emphasis weights. The weighting mechanism is computationally light – roughly two orders of magnitude cheaper than typical attention – while still adapting to diverse inputs.
Hierarchical Processing The system uses a tiered routing strategy:
Cache tier – Patterns seen before are resolved instantly
Standard tier – New confident cases proceed through the full pipeline
Deep tier – Ambiguous cases receive additional scrutiny (rare)
Results from deeper tiers feed back into the cache, so the system becomes faster over time.
Prediction & Explanation A compact ensemble produces the final output. The system returns the prediction, a confidence score, the top contributing factors, and – if data was missing – specific recommendations to improve future predictions.
Continuous Monitoring An online detector watches for changes in prediction patterns. When behaviour shifts, the system re‑calibrates automatically in seconds, not weeks.
Current Limitations – Honest and Clear
Synthetic validation only – real‑world public datasets (creditcard.csv, MIMIC‑III, CMAPSS) are currently being tested.
Not for unstructured data – images, free text, video are out of scope.
Not a replacement for LLMs – it is a specialised tool for structured data.
Inference speed measured under controlled conditions – real‑world network latency may add overhead.
Why This Matters Now
The AI industry is shifting. The conversation is no longer just about scale – it is about sustainability, accessibility, and trust.
95% of GenAI pilots fail because LLMs cannot reliably handle structured business data.
Edge AI requires models that run on cheap, low‑power hardware.
Regulations (EU AI Act, etc.) demand explainability for high‑stakes decisions (credit, healthcare, public safety).
TinyBrain++ is a working example of a different direction: compact, efficient, transparent, and deployable anywhere.