Antaḥkaraṇa-Net
🔗 Model & code: 🤗 huggingface.co/deepakdsoni/antahkarana-v1 · github.com/deepaksatna/antahkarana
A working AI architecture built on the 2,500-year-old Vedic / Sanskrit model of mind — the antaḥkaraṇa ("inner instrument"). One agent that learns continually without forgetting, scales its effort and mood to its own state, perceives without hallucinating, runs embodied and even on a spiking (neuromorphic-style) substrate — all validated on real hardware. A deep-research proof-of-concept: the foundation is built and measured honestly; scaling it is the next chapter.
📦 Model family
| Model | What |
|---|---|
| antahkarana-v1 | the original architecture + v1 vision models — the most stable continual learner (only positive backward transfer) |
| antahkarana-v2 | accuracy-recovering v2 (36.5M) — matches SOTA accuracy at ~3× less forgetting |
| antahkarana-7B | the architecture scaled to a 7B language model |
🔬 Scale-Up Benchmark & Models
The original POC is now scaled to real WideResNets (36.5M–52.6M params) and benchmarked by live inference on the trained checkpoints — 11 models, 7 capabilities each, on a single NVIDIA A10. Full report with methodology and per-model tables: BENCHMARK_REPORT.md. End-to-end live test on the published model: E2E_TEST_REPORT.md (run it: python3 e2e_demo.py).
Models in this repo
| File | Params | Dataset / setup | Forgetting ↓ |
|---|---|---|---|
antahkarana-36.5M-cifar100-wrn28-10.pt |
36.5M | CIFAR-100, 10-task (primary) | 0.565→0.018 (31.8×) |
antahkarana-52.6M-cifar100-wrn28-12.pt |
52.6M | CIFAR-100, 10-task | 0.542→0.021 (25.4×) |
antahkarana-36.5M-tinyimagenet-wrn28-10.pt |
36.5M | Tiny-ImageNet, 10-task | 0.503→0.017 (29.2×) |
antahkarana-36.5M-cifar100-20task-wrn28-10.pt |
36.5M | CIFAR-100, 20-task lifelong | 0.603→0.049 (12.3×) |
Load any with the self-contained load_akn.py (only needs PyTorch):
from load_akn import load
model, ck = load("antahkarana-36.5M-cifar100-wrn28-10.pt")
logits = model(x, task=0) # x: (N, 3, 32, 32)
print(ck["results"]["agent"])
Catastrophic forgetting — cut 12–41× vs a naive baseline
Two model sizes, two datasets, 10- and 20-task streams. The agent forgets almost nothing (0.01–0.05); the naive net collapses to its last task.
Capability scorecard — 10 of 11 models pass 7/7
The single 6/7 (āśrama_s0) is a borderline threshold artifact on avg-accuracy (0.592 vs the 0.60 = 3×-chance bar); its memory/abstention/calibration all pass.
Per-task retention — nothing collapses
Final accuracy of every task after the full stream (10-task and 20-task). All stay well above chance — the model remembers task 0 after learning task 19.
Anti-hallucination (pramāṇa) — abstains instead of guessing
Calibrated abstention: gated accuracy 0.91–1.00 when it commits; abstains on up to 99.7% of out-of-distribution (SVHN) inputs.
Legible mind-state — plasticity falls across the four āśrama life-stages
The 20-task lifelong run: plasticity headroom drops childhood→old-age (brāhmacarya→gṛhastha→vānaprastha→saṃnyāsa) while forgetting stays low and bounded.
Headline (5-/2-seed means): forgetting 0.589→0.0146 (~41×) at 36.5M; 27× at 52.6M; 29× on Tiny-ImageNet; 12.5× on the 20-task lifelong run. Gated accuracy 0.93–0.96, calibration ECE cut ~5–7×, OOD abstention up to 99.7% — all by live inference on the released checkpoints.
1. What it is, in one breath
Modern AI already has the pieces of a mind — attention, memory, decision, control — but no principled way to wire them into one self-regulating, lifelong-learning whole. The Vedic model of mind is exactly such a wiring diagram. Antaḥkaraṇa-Net implements it: every Sanskrit faculty becomes a real ML module, assembled into a single agent.
| Sanskrit faculty | What it does | ML module |
|---|---|---|
| manas (मनस्) | attention / perception gate | precision-weighted attention encoder |
| buddhi (बुद्धि) | discrimination, decision | evidence-accumulation / executive |
| ahaṃkāra (अहंकार) | the "I-maker" / self-model | identity latent |
| chitta (चित्त) | memory & the subconscious | continual memory (EWC + decay) |
| guṇas (सत्त्व·रजस्·तमस्) | the three qualities | one controller of plasticity / explore / consolidate |
| tapas (तपस्) | concentrated effort | effort-allocation by need |
| divya-dṛṣṭi / pramāṇa | valid extended perception | calibrated abstention gate (anti-hallucination) |
| turīya (तुरीय) | the witness | reward-invariant identity monitor |
| āśrama | life-stages | lifelong plasticity schedule (childhood → old age) |
2. Why — the motivation
Two motivations meet here.
(a) The Vedic psychology is a stunningly good systems diagram of mind. The Upaniṣads, Sāṃkhya and
Yoga decompose cognition into a four-fold inner instrument, separate awareness from processing (the
"hard problem", 2,000 years early), give a four-state model of consciousness (waking / dream / deep-sleep
/ turīya), and a real theory of the subconscious (saṃskāra / vāsanā). It even contains a developmental
law — the āśramas — for how a mind should keep improving across a whole lifetime. (The full study is in
philosophy/: texts & mantras, the modern-neuroscience cross-walk, the Sanskrit formulae,
the modern equations, and the architecture derivation.)
(b) Today's AI has matching blind spots — and the Vedic model addresses each one:
| Limitation of today's models | Antaḥkaraṇa-Net's structural answer |
|---|---|
| No continual learning (frozen after training) | chitta: incremental updates, never retrain from zero |
| Catastrophic forgetting | saṃskāra importance with growth and decay |
| Full-retrain energy cost (GWh) | update + sleep-time consolidation; spiking substrate |
| Dense, always-full-power compute | guṇa-scaled effort; event-driven spikes |
| No self-monitoring across modes | turīya reward-invariant monitor |
| Confident hallucination | pramāṇa validity gate (abstain, don't confabulate) |
| Disembodied (no action→consequence) | karma loop + battery→guṇa in the embodied agent |
3. How it works — the architecture
The backbone thinks; the Vedic layer remembers, regulates, perceives, and monitors around it:
- Perception → decision pipeline: manas (attention) → chitta (memory) → buddhi (decision) ↔ ahaṃkāra (self).
- One guṇa controller turns a 3-vector (sattva, rajas, tamas) into all the learning dynamics (plasticity, exploration, consolidation, pruning) — and it is forgetting-aware (protect hard tasks, back off on easy ones) and, when embodied, driven by the battery (low battery → tamas → conserve).
- Four operating states: jāgrat (wake/act) → svapna (dream/replay) → suṣupti (sleep/consolidate).
- Two safety overlays: the pramāṇa gate (extended perception must be valid knowledge, not fancy) and the turīya witness (a reward-invariant identity monitor).
- The āśrama schedule keeps plasticity non-zero for life and re-opens critical periods on novelty — the "always enhanceable, childhood → old age" property.
Because the control layer is backbone-agnostic, the same agent runs on a toy MLP, a real CNN, an RL policy, or a spiking net — which is why embodiment and neuromorphic are extensions of one model, not separate builds.
4. What we achieved — results (all from real runs)
| Phase | Result | Status |
|---|---|---|
| Integration | one agent, all faculties in a single wake/dream/sleep loop | ✅ |
| Phase II — real CNN + data (GPU) | Split-CIFAR: forgetting 0.217 → 0.008 (~27×); +9–10 pts accuracy | ✅ |
| Forgetting-aware controller | fixed the easy-task over-regularization (MNIST 0.974 → 0.986), kept the CIFAR win | ✅ |
| Continual benchmark | consolidation cuts forgetting ~60–80× (0.242 → 0.003) | ✅ |
| Divya-dṛṣṭi + Pramāṇa | accepted-prediction accuracy rises 0.80 → 0.91, abstains on blind inputs | ✅ |
| Track B — embodiment | karma loop (success 1.00 vs random 0.30); battery→guṇa (ε 0.087 hungry → 0.122 charged); retention across 4 regimes 0.38 → 1.00 | ✅ |
| Track C — neuromorphic (spiking) | the spiking net works — matches ANN accuracy (0.943 vs 0.929) at 10.7% spike density; conservative ~1.9× software energy floor | ✅ spiking proven · ⏳ only chip deployment pending |
About Track C "pending". The spiking network is done and working — it runs and matches the normal network's accuracy, which proves the architecture runs on event-driven (neuromorphic-style) computation. What is pending is only deployment to a real neuromorphic chip (Intel Loihi 2 / BrainChip Akida), which we don't have. The ~1.9× is a deliberately conservative software estimate (per-operation energy only); the famous 100–1000× neuromorphic figures come from chip-only effects (event-skipping, in-memory compute, no data movement) that a GPU simulation cannot reproduce — so we report the floor, not the headline. Un-pending it needs a neuromorphic board + a port via Intel Lava; it is the only step in the whole project gated on hardware rather than code.
Full numbers, seeds, and the honest caveats are in RESULTS.md; the staged plan is in
ROADMAP.md.
5. Why it's different (and why that matters)
- Most "continual learning" papers fix one mechanism. This is a single agent that unifies memory, effort, control, perception-validity and self-monitoring under one interpretable scheme — with an observable "mind-state" trace (life-stage, guṇa mix, plasticity, witness drift) you can read as it lives.
- It learns forever without forgetting — and the controller learns when to protect, so it doesn't over-regularize easy tasks (a failure mode we caught and fixed honestly).
- It is honest about hallucination. The pramāṇa gate is the engineering form of the Nyāya rule that extraordinary perception must be a valid means of knowledge — it abstains rather than confabulate.
- It carries from supervised → embodied → spiking unchanged, because the architecture (not a trick) is the contribution.
- It is a research POC, stated plainly. Strong faculties (consolidation, replay, pramāṇa,
forgetting-aware control, task-conditioned policy) are proven; modest ones (tapas, decay) and conceptual
ones (āśrama, the witness) are labeled as such — see the component scorecard in
RESULTS.md.
6. How it advances current AI research
A research proof-of-concept — but one that speaks directly to several of the field's most active open problems, offering a principled, reproducible framework rather than a point fix.
Continual / lifelong learning. Catastrophic forgetting is one of ML's hardest open problems — today's large models are effectively frozen after training and must be expensively re-trained to absorb new knowledge. This unifies importance-based consolidation, rehearsal, and a forgetting-aware controller into a single agent that learns indefinitely, cutting forgetting ~6–80× in our runs.
Compute & energy sustainability. Frontier-model (re)training consumes gigawatt-hours. The architecture offers two complementary levers — incremental updates (no retrain-from-scratch) and an event-driven spiking path (validated in software, matching ANN accuracy) — a concrete route toward order-of-magnitude lower inference energy on neuromorphic hardware.
Reliability & hallucination. Models routinely assert what they don't know. The Pramāṇa validity gate provides calibrated abstention — "know when you don't know" — a deployable anti-hallucination primitive grounded in epistemology (extended perception must be a valid means of knowledge, not fancy).
Safety & alignment. The turīya reward-invariant monitor, plus the siddhi principle — capabilities must stay subordinate to the goal (Yoga Sūtra 3.37) — anticipate modern instrumental-goal / mesa-optimization concerns and provide a structural oversight pattern, not an afterthought.
Self-regulation & adaptivity. A single interpretable guṇa signal auto-balances exploration, consolidation, and conservation; the controller learns when to protect (fixing over-regularization on easy tasks), and — when embodied — ties learning dynamics to real resource state (battery → guṇa).
Interpretability. Unlike opaque agents, it exposes a readable "mind-state" trajectory — life-stage, guṇa mix, plasticity headroom, identity drift — making the learning process auditable.
Embodied & agentic AI. The karma loop (action → consequence → disposition) and metabolic-state-driven behavior give a principled scaffold for autonomous agents that learn and self-regulate in the world, not just on a dataset.
A bridge from cognitive science to ML. Rather than ad-hoc tricks, it contributes a coherent, theory-grounded cognitive architecture — a template for composing modular faculties into one self-regulating whole, shipped as an open, modular library (ChittaKit) that drops into any PyTorch backbone, with a transparent, falsifiable results scorecard others can build on.
In short: it reframes a set of disconnected AI problems — forgetting, energy, hallucination, alignment, adaptivity, interpretability — as facets of one missing capability: a principled architecture for a mind that learns for life and regulates itself. That reframing, with working evidence, is the contribution.
7. Run it
pip install -r requirements.txt # torch, numpy (+ torchvision/snntorch for Phase II/C)
cd experiments
python3 integrated_agent.py # ★ the whole agent + its mind-state trace (CPU, ~1 min)
python3 capacity_benchmark.py # consolidation / decay / tapas ablation
python3 divya_drsti.py # extended perception + Pramāṇa validity gate
python3 track_b_v2.py # embodiment: karma loop + battery→guṇa + retention
# GPU (e.g. CUDA_VISIBLE_DEVICES=0):
python3 phase2_vision.py --dataset cifar10 # real CNN on Split-CIFAR
python3 track_c_spiking.py # spiking perception net (snnTorch)
python3 make_figures.py # regenerate the README figures
chittakit/ the novel modules — saṃskāra · guṇa · meta-guṇa · āśrama · tapas · pramāṇa · witness · antahkarana
experiments/ integrated_agent · capacity/continual benchmarks · divya_drsti · sanjaya · track_b · track_c · phase2_vision
philosophy/ the deep study: texts & mantras → modern science → Sanskrit formulae → math models → architecture
assets/ banner · architecture diagram · results figures
RESULTS.md every number + the honest scorecard ROADMAP.md what's done / what's next
8. How it can be extended (it's amazing because it can grow)
- Scale the backbone — drop in a ResNet/ViT or a transformer; the Vedic layer is unchanged.
- Neuromorphic hardware — map the spiking parts to Loihi 2 / Akida (via Intel Lava) for the milliwatt, always-on form — the full energy thesis.
- Real robot — the embodied agent → Jetson + ROS 2, with a real battery driving the guṇa and an event camera feeding manas.
- Meta-learn the controllers — guṇa and tapas via meta-gradient / population-based training.
- Richer sims — MuJoCo / Isaac for physics; a continual RL stream for the karma loop at scale.
Each is extension, not invention: the hard part — assembling a coherent, lifelong, self-regulating agent from the Vedic model and proving it on real and spiking hardware — is done.
9. Honest scope
This is a deep-research proof-of-concept at modest scale (small CNNs, MNIST/CIFAR, a gridworld) — enough to prove the architecture works, not to rival a frontier model. Every negative result (the v1 RL-retention miss, the neutral evolution-strategy run, the MNIST over-regularization) was diagnosed and either fixed or kept as a labeled limitation — the discipline that makes the positive results trustworthy. The Vedic↔ML mappings are engineering analogies, clearly flagged; no claim is made that the texts contain neuroscience, and nothing here is conscious.
Code: MIT. Built on the Upaniṣads, Sāṃkhya, Yoga, and modern ML (PyTorch · snnTorch). Part of a deep study
of the Vedic philosophy of mind — see philosophy/.