File size: 17,221 Bytes
ced4fa3
 
b585e18
 
7ac0c24
 
 
 
b585e18
 
 
 
 
ced4fa3
25345f7
7ac0c24
b585e18
7ac0c24
25345f7
7ac0c24
b585e18
 
25345f7
b585e18
 
 
 
 
 
 
 
 
 
 
7ac0c24
b585e18
 
7ac0c24
b585e18
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
25345f7
b585e18
7ac0c24
 
b585e18
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
7ac0c24
b585e18
 
 
 
 
 
 
 
7ac0c24
 
 
b585e18
 
 
 
 
 
 
 
 
 
 
 
7ac0c24
b585e18
 
7ac0c24
ced4fa3
b585e18
 
 
 
 
 
 
 
ced4fa3
7ac0c24
b585e18
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
ced4fa3
b585e18
 
 
 
7ac0c24
b585e18
 
7ac0c24
b585e18
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
7ac0c24
 
b585e18
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
7ac0c24
b585e18
 
 
 
 
 
 
7ac0c24
b585e18
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
ced4fa3
b585e18
7ac0c24
b585e18
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
18a9623
7ac0c24
 
25345f7
b585e18
 
 
 
7ac0c24
b585e18
 
 
 
 
 
ced4fa3
b585e18
 
 
 
 
 
 
 
 
 
7ac0c24
b585e18
7ac0c24
18a9623
b585e18
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
7ac0c24
b585e18
 
 
ced4fa3
18a9623
7ac0c24
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
import gradio as gr
import torch
import torch.nn as nn
import torch.optim as optim
from transformers import AutoTokenizer, AutoModel
from sklearn.metrics.pairwise import cosine_similarity
import numpy as np
from collections import deque
import random
from scipy.stats import entropy
import matplotlib.pyplot as plt
import pandas as pd
from sklearn.manifold import TSNE


tokenizer = AutoTokenizer.from_pretrained("sentence-transformers/all-MiniLM-L6-v2")
base_model = AutoModel.from_pretrained("sentence-transformers/all-MiniLM-L6-v2")


device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
base_model.to(device)


class ProjectionHead(nn.Module):
    def __init__(self, input_dim=384, hidden_dim=128, output_dim=384):
        super().__init__()
        self.projection = nn.Sequential(
            nn.Linear(input_dim, hidden_dim),
            nn.ReLU(),
            nn.Linear(hidden_dim, output_dim)
        )
        
    def forward(self, x):
        return self.projection(x)

projection_head = ProjectionHead().to(device)
optimizer = optim.Adam(projection_head.parameters(), lr=0.001)

# Hierarchical concept structure
class ConceptHierarchy:
    def __init__(self):
        self.hierarchy = {
            "health": ["physical", "mental", "holistic", "preventive"],
            "tech": ["software", "hardware", "AI", "blockchain"],
            "nature": ["ecology", "wildlife", "climate", "conservation"],
            "spirit": ["mindfulness", "philosophy", "religion", "consciousness"]
        }
        self.reverse_lookup = {}
        for main, subs in self.hierarchy.items():
            for sub in subs:
                self.reverse_lookup[sub] = main
    
    def get_parent(self, subcategory):
        return self.reverse_lookup.get(subcategory, subcategory)
    
    def get_children(self, category):
        return self.hierarchy.get(category, [])
    
    def all_categories(self):
        all_cats = list(self.hierarchy.keys())
        for subs in self.hierarchy.values():
            all_cats.extend(subs)
        return all_cats

concept_hierarchy = ConceptHierarchy()


class CognitiveMemory:
    def __init__(self, max_length=100):
        self.samples = deque(maxlen=max_length)
        self.embeddings_cache = {}
        self.concept_centroids = {}
        self.uncertainty_history = []
        self.drift_scores = {}
        
    def add(self, text, label, embedding=None):
        if embedding is None:
            embedding = embed_text(text)
        
        self.samples.append((text, label, embedding))
        
        # Update concept centroids
        if label not in self.concept_centroids:
            self.concept_centroids[label] = embedding
        else:
            # Moving average update
            self.concept_centroids[label] = 0.9 * self.concept_centroids[label] + 0.1 * embedding
            
        # Check for concept drift
        if len(self.samples) > 10:
            self._detect_concept_drift()
    
    def _detect_concept_drift(self):
        # Simple drift detection by measuring distance change over time
        for label in self.concept_centroids:
            recent_examples = [emb for txt, lbl, emb in self.samples if lbl == label][-5:]
            if len(recent_examples) > 1:
                recent_centroid = torch.stack(recent_examples).mean(dim=0)
                drift = torch.norm(self.concept_centroids[label] - recent_centroid).item()
                self.drift_scores[label] = drift
    
    def get_embeddings_labels(self):
        if not self.samples:
            return None, None, None
        texts, labels, embeddings = zip(*self.samples)
        return embeddings, labels, texts
    
    def get_drift_report(self):
        if not self.drift_scores:
            return "No drift detected yet"
        
        highest_drift = max(self.drift_scores.items(), key=lambda x: x[1])
        if highest_drift[1] > 0.15:
            return f"Significant concept drift detected in '{highest_drift[0]}' category"
        return "Concept stability maintained across all categories"

# Enhanced embedding with adaptive projection
def embed_text(text, apply_projection=False):
    inputs = tokenizer(text, return_tensors="pt", truncation=True, padding=True).to(device)
    with torch.no_grad():
        outputs = base_model(**inputs)
        embedding = outputs.last_hidden_state.mean(dim=1).squeeze().cpu()
    
    if apply_projection:
        with torch.no_grad():
            embedding = projection_head(embedding.to(device)).cpu()
    
    return embedding

# Initialize memory
memory = CognitiveMemory()

# Contrastive learning update
def update_projection_head(pos_examples, neg_examples, temperature=0.1):
    projection_head.train()
    optimizer.zero_grad()
    
    # Prepare positive and negative examples
    pos_embeddings = torch.stack([ex.to(device) for ex in pos_examples])
    neg_embeddings = torch.stack([ex.to(device) for ex in neg_examples])
    
    # Project embeddings
    pos_projections = projection_head(pos_embeddings)
    neg_projections = projection_head(neg_embeddings)
    
    # Calculate similarities
    pos_sim = torch.mm(pos_projections, pos_projections.t()) / temperature
    neg_sim = torch.mm(pos_projections, neg_projections.t()) / temperature
    
    # Create contrastive loss
    logits = torch.cat([pos_sim, neg_sim], dim=1)
    labels = torch.arange(pos_projections.size(0)).to(device)
    
    # Loss calculation (simplified contrastive loss)
    loss = nn.CrossEntropyLoss()(logits, labels)
    loss.backward()
    optimizer.step()
    
    projection_head.eval()
    return loss.item()

# Active learning sample selection
def get_informative_samples(embeddings, labels, num_samples=3):
    if len(set(labels)) < 2:
        return ["Need examples from multiple categories"]
    
    # Calculate uncertainty for each category
    label_set = set(labels)
    uncertainty_scores = {}
    
    for category in label_set:
        # Get other category embeddings
        other_embeds = [e for e, l in zip(embeddings, labels) if l != category]
        if not other_embeds:
            continue
        
        # Calculate centroid for this category
        this_embeds = [e for e, l in zip(embeddings, labels) if l == category]
        centroid = torch.stack(this_embeds).mean(dim=0)
        
        # Calculate similarity to other categories
        other_stack = torch.stack(other_embeds)
        sims = torch.matmul(centroid.unsqueeze(0), other_stack.transpose(0, 1))
        
        # Higher max similarity means more ambiguity/uncertainty
        uncertainty_scores[category] = torch.max(sims).item()
    
    # Find the most uncertain categories
    sorted_categories = sorted(uncertainty_scores.items(), key=lambda x: -x[1])
    
    # Suggest example prompts for the most uncertain categories
    suggestions = []
    for category, score in sorted_categories[:2]:
        subcategories = concept_hierarchy.get_children(category)
        if subcategories:
            suggestions.append(f"Need examples distinguishing '{category}' from other categories")
            suggestions.append(f"Consider examples about '{random.choice(subcategories)}'")
    
    return suggestions

# Uncertainty quantification
def calculate_uncertainty(similarities):
    # Convert to probability distribution
    probs = similarities / np.sum(similarities)
    
    # Calculate entropy (higher means more uncertain)
    uncertainty = entropy(probs)
    
    # Normalize between 0 and 1
    max_entropy = np.log(len(probs))
    normalized_uncertainty = uncertainty / max_entropy if max_entropy > 0 else 0
    
    return normalized_uncertainty

# Counterfactual explanation generation
def generate_counterfactual(text_embedding, predicted_label, labels, embeddings):
    # Find nearest example of a different class
    different_class_embeddings = [(e, l, i) for i, (e, l) in enumerate(zip(embeddings, labels)) if l != predicted_label]
    
    if not different_class_embeddings:
        return "No alternative classes available for counterfactual"
    
    # Calculate distances
    distances = [torch.norm(text_embedding - e).item() for e, _, _ in different_class_embeddings]
    nearest_idx = np.argmin(distances)
    nearest_embed, nearest_label, original_idx = different_class_embeddings[nearest_idx]
    
    # Calculate direction vector to move from current to alternate class
    direction = nearest_embed - text_embedding
    direction_normalized = direction / torch.norm(direction)
    
    # Identify key dimensions (simplified)
    key_dims = torch.topk(torch.abs(direction_normalized), 10).indices
    
    return f"To change classification from '{predicted_label}' to '{nearest_label}', the text would need more emphasis on concepts found in '{labels[original_idx]}'"

# Advanced inference with uncertainty and counterfactuals
def infer_with_insights(text):
    if len(memory.samples) < 5:
        return "Label: Unknown", "Insight: Need more training examples (at least 5)", "Uncertainty: High", "Visualization not available", "No counterfactual available"
    
    # Get text embedding
    input_embedding = embed_text(text, apply_projection=True)
    
    # Get memory contents
    memory_embeddings, memory_labels, memory_texts = memory.get_embeddings_labels()
    memory_embeddings = [embed_text(mem_text, apply_projection=True) for mem_text in memory_texts]
    
    # Calculate similarities
    input_vec_np = input_embedding.unsqueeze(0).numpy()
    memory_vecs_np = torch.stack(memory_embeddings).numpy()
    sims = cosine_similarity(input_vec_np, memory_vecs_np)[0]
    
    # Find best match
    best_idx = np.argmax(sims)
    confidence = sims[best_idx]
    predicted_label = memory_labels[best_idx]
    
    # Calculate uncertainty
    uncertainty = calculate_uncertainty(sims)
    uncertainty_level = "High" if uncertainty > 0.8 else "Medium" if uncertainty > 0.5 else "Low"
    
    # Generate counterfactual
    counterfactual = generate_counterfactual(input_embedding, predicted_label, memory_labels, memory_embeddings)
    
    # Generate hierarchical insight
    parent_category = concept_hierarchy.get_parent(predicted_label)
    subcategories = concept_hierarchy.get_children(parent_category)
    
    if predicted_label in subcategories:
        insight = f"This concept falls under '{parent_category}' with specific focus on '{predicted_label}' aspects."
    else:
        subcategory_text = ", ".join(subcategories[:2]) + ("..." if len(subcategories) > 2 else "")
        insight = f"This concept broadly relates to '{predicted_label}' which includes aspects like {subcategory_text}."
    
    # Create visualization data
    tsne = TSNE(n_components=2, random_state=42)
    all_embeddings = memory_vecs_np.tolist() + [input_vec_np[0].tolist()]
    all_labels = list(memory_labels) + ["Current Input"]
    
    # Create visualization code
    vis_code = """
    ```python
    # Load this code in a notebook to visualize
    import matplotlib.pyplot as plt
    import numpy as np
    from sklearn.manifold import TSNE
    
    # Your embeddings and labels would go here
    # This is a placeholder visualization
    plt.figure(figsize=(10, 8))
    for label in set(labels[:-1]):
        indices = [i for i, l in enumerate(labels[:-1]) if l == label]
        plt.scatter(coords[indices, 0], coords[indices, 1], label=label)
    
    # Highlight the input point
    plt.scatter(coords[-1, 0], coords[-1, 1], color='red', 
                s=100, marker='*', label='Current Input')
    
    plt.legend()
    plt.title("Concept Map Visualization")
    plt.show()
    ```
    """
    
    # Add uncertainty and drift detection to memory
    memory.uncertainty_history.append(uncertainty)
    
    drift_report = memory.get_drift_report()
    full_insight = f"{insight}\n\n{drift_report}"
    
    return f"Label: {predicted_label} (Confidence: {confidence:.2f})", full_insight, f"Uncertainty: {uncertainty_level} ({uncertainty:.2f})", vis_code, counterfactual

# Enhanced training with contrastive learning
def train_sample(text, label):
    # Check if we have enough samples for contrastive learning
    embeddings, labels, _ = memory.get_embeddings_labels() or ([], [], [])
    
    text_embedding = embed_text(text)
    
    # Add to memory
    memory.add(text, label, embedding=text_embedding)
    
    # If we have multiple categories, do contrastive update
    unique_labels = set(labels) if labels else set()
    if label in unique_labels and len(unique_labels) > 1:
        # Get positive examples (same label)
        pos_examples = [e for e, l in zip(embeddings, labels) if l == label]
        
        # Get negative examples (different labels)
        neg_examples = [e for e, l in zip(embeddings, labels) if l != label]
        
        # If we have enough examples, do contrastive update
        if len(pos_examples) > 0 and len(neg_examples) > 0:
            loss = update_projection_head(
                pos_examples[:min(5, len(pos_examples))], 
                neg_examples[:min(5, len(neg_examples))]
            )
            contrastive_msg = f" • Updated adaptive projection (loss: {loss:.4f})"
        else:
            contrastive_msg = ""
    else:
        contrastive_msg = ""
    
    # Get active learning suggestions if we have enough samples
    if len(memory.samples) >= 5:
        active_suggestions = get_informative_samples(embeddings + [text_embedding], labels + [label])
        active_msg = "\n\nSuggested next examples:\n" + "\n".join([f"• {s}" for s in active_suggestions])
    else:
        active_msg = "\n\nAdd " + str(5 - len(memory.samples)) + " more examples to enable active learning."
    
    return f"Stored '{text}' as '{label}' | Total samples: {len(memory.samples)}{contrastive_msg}{active_msg}"

# Gradio UI
with gr.Blocks() as app:
    gr.Markdown("# Vers3Dynamics Labeling System")
    gr.Markdown("### This system features meta-learning, active learning, uncertainty quantification, and concept drift detection")
    
    with gr.Row():
        text_input = gr.Textbox(label="Input Text", placeholder="Type a concept like 'Blockchain for healthcare records'...")
    
    infer_btn = gr.Button("Analyze with Cognitive Insights")
    
    with gr.Row():
        label_output = gr.Textbox(label="Classification Result")
        insight_output = gr.Textbox(label="Cognitive Insight")
        
    with gr.Row():
        uncertainty_output = gr.Textbox(label="Uncertainty Analysis")
        counterfactual_output = gr.Textbox(label="Counterfactual Explanation")
    
    visualization_output = gr.Code(label="Visualization Code", language="python")
    
    infer_btn.click(
        fn=infer_with_insights, 
        inputs=text_input, 
        outputs=[label_output, insight_output, uncertainty_output, visualization_output, counterfactual_output]
    )
    
    gr.Markdown("### Cognitive Training")
    
    with gr.Row():
        train_text = gr.Textbox(label="Training Example")
        
    with gr.Row():
        main_categories = gr.Radio(list(concept_hierarchy.hierarchy.keys()), label="Main Category")
        sub_categories = gr.Dropdown([], label="Sub-Category (Optional)")
    
    def update_subcategories(main_category):
        if main_category:
            return gr.Dropdown.update(choices=[""] + concept_hierarchy.get_children(main_category))
        return gr.Dropdown.update(choices=[])
    
    main_categories.change(fn=update_subcategories, inputs=main_categories, outputs=sub_categories)
    
    train_btn = gr.Button("Store & Learn From Example")
    train_output = gr.Textbox(label="Training Status & Suggestions")
    
    def handle_training(text, main_category, sub_category):
        # Use subcategory if provided, otherwise use main category
        final_category = sub_category if sub_category else main_category
        return train_sample(text, final_category)
    
    train_btn.click(
        fn=handle_training, 
        inputs=[train_text, main_categories, sub_categories], 
        outputs=train_output
    )
    
    # System status section
    gr.Markdown("### Vers3Dynamics System Status")
    
    def get_system_status():
        if len(memory.samples) == 0:
            return "System initialized - no training data yet"
        
        num_samples = len(memory.samples)
        _, labels, _ = memory.get_embeddings_labels()
        category_counts = {}
        for label in labels:
            if label in category_counts:
                category_counts[label] += 1
            else:
                category_counts[label] = 1
        
        categories_info = ", ".join([f"{k}: {v}" for k, v in category_counts.items()])
        
        adaptations = "Meta-learning projection: " + ("Active" if len(memory.samples) > 5 else "Not yet active")
        
        drift_info = memory.get_drift_report()
        
        return f"System Status:\n• Samples: {num_samples}\n• Categories: {categories_info}\n• {adaptations}\n• {drift_info}"
    
    status_btn = gr.Button("Check System Status")
    status_output = gr.Textbox(label="Current System Status")
    status_btn.click(fn=get_system_status, outputs=status_output)

if __name__ == "__main__":
    app.launch()