Are Candidate Models Really Needed for Active Learning?

Paper: Are Candidate Models Really Needed for Active Learning? Harshini Mridula Mohan, Maanya Manjunath, Vipul Arya, S.H. Shabbeer Basha, Nitin Cheekatla Preprint submitted to Computer Vision and Image Understanding, May 2026

This repository contains the official implementation of our Deep Active Learning (DAL) framework. We demonstrate that models with randomly initialized weights can achieve competitive or superior performance compared to active learning methods that rely on pre-trained candidate models — eliminating the computational overhead of candidate model training entirely.

We evaluate three confidence-based sampling strategies:

• HC — High Confidence: selects samples the model is most certain about
• LC — Low Confidence: selects samples the model is most uncertain about
• HCLC — High Confidence initially, then Low Confidence in subsequent rounds

---

Results

Table 1: CIFAR-10

Model	Method	Time Saved (hrs)	Accuracy (%)
DenseNet-121	LC (10K)	0.16	91.37 ± 0.20
	LC	0.16	92.87 ± 0.07
	HC	0.16	92.72 ± 0.16
	HCLC	0.16	93.08 ± 0.23
ResNet-56	LC	0.26	91.60 ± 0.12
	HC	0.26	91.30 ± 0.09
	HCLC	0.26	91.60 ± 0.12
VGG-16	LC (20K)	0.5	84.26
	LC (40K)	0.5	91.89
	LC	0.5	94.21 ± 0.14
	HC	0.5	93.94 ± 0.19
	HCLC	0.5	94.20 ± 0.11
ResNet-18	LC (5K)	0.25	81.27 ± 0.12
	LC (10K)	0.25	90.12 ± 0.07
	LC (40K)	0.25	92.69 ± 0.09
	LC	0.25	93.53 ± 0.08
	HC	0.25	93.28 ± 0.03
	HCLC	0.25	93.48 ± 0.12
Swin Transformer	LC	0.06	86.23 ± 0.10
	HC	0.06	83.88 ± 0.35
	HCLC	0.06	85.80 ± 0.02
ViT-Small	LC	0.52	83.92 ± 0.36
	HC	0.52	82.70 ± 0.18
	HCLC	0.52	83.81 ± 0.21
MobileNetV2	LC (10K)	0.52	82.53
	LC	0.20	94.16 ± 0.03
	HC	0.20	92.84 ± 0.23
	HCLC	0.20	94.16 ± 0.03

---

Table 2: CIFAR-100

Model	Method	Time Saved (hrs)	Accuracy (%)
DenseNet-121	LC (10K)	0.19	59.03
	LC	0.19	71.65 ± 0.15
	HC	0.19	70.98 ± 0.19
	HCLC	0.19	71.33 ± 0.24
ResNet-56	LC	0.31	66.22 ± 0.21
	HC	0.31	66.30 ± 0.47
	HCLC	0.31	66.39 ± 0.06
VGG-16	LC	0.57	66.46 ± 0.63
	HC	0.57	64.87 ± 0.35
	HCLC	0.57	66.11 ± 0.16
ResNet-18	LC (10K)	0.10	59.01 ± 0.22
	LC	0.10	73.24 ± 0.19
	HC	0.10	71.97 ± 0.05
	HCLC	0.10	73.42 ± 0.24
MobileNetV2	LC	0.25	73.79 ± 0.13
	HC	0.25	72.82 ± 0.13
	HCLC	0.25	73.79 ± 0.13

---

Table 3: SVHN

Model	Method	Time Saved (hrs)	Accuracy (%)
DenseNet-121	LC	0.94	95.77 ± 0.01
	HC	0.94	95.48 ± 0.15
	HCLC	0.94	95.76 ± 0.08
ResNet-56	LC	0.35	96.12 ± 0.11
	HC	0.35	95.99 ± 0.05
	HCLC	0.35	96.12 ± 0.11
VGG-16	LC (50K)	0.27	94.22
	LC	0.27	95.51 ± 0.08
	HC	0.27	95.45 ± 0.07
	HCLC	0.27	95.61 ± 0.09
ResNet-18	LC (15K)	0.29	91.80 ± 0.06
	LC (50K)	0.29	93.23 ± 0.30
	LC	0.29	95.84 ± 0.09
	HC	0.29	95.65 ± 0.08
	HCLC	0.29	95.83 ± 0.02

---

Table 4: TinyImageNet (ResNet-18)

Method	Time Saved (hrs)	Annotation Sim. Time (hrs)	Accuracy (%)
LC	1.20	29.87	55.99 ± 0.12
HC	1.20	29.87	54.94 ± 0.14
HCLC	1.20	29.87	55.92 ± 0.11

BADGE took ~45 hrs — 1.5× longer than our methods — for only a marginal accuracy gain of 0.74%.

---

Table 5: PASCAL VOC 2012 — Object Detection (SSD / VGG-16)

Method	SSD Variant	Time Saved (hrs)	mAP
LC	SSD / VGG-16	2.58	81.53 ± 0.09
HC	SSD / VGG-16	2.58	80.97 ± 0.21
HCLC	SSD / VGG-16	2.58	78.13 ± 0.09

---

Table 6: Ablation Study — CIFAR-10 (Different Acquisition Function Combinations)

Model	Method	Accuracy (%)
DenseNet-121	LCHC	94.49
	HLH (hybrid)	94.52
	RHC	94.47
	RLC	94.35
ResNet-56	LCHC	92.99
	HLH (hybrid)	93.41
	RHC	91.89
	RLC	92.30
VGG-16	LCHC	93.97
	HLH (hybrid)	94.35
	RHC	94.05
	RLC	94.17
ResNet-18	LCHC	95.24
	HLH (hybrid)	95.62
	RHC	95.24
	RLC	95.64
MobileNetV2	LCHC	94.67
	HLH (hybrid)	95.60
	RHC	94.69
	RLC	95.61

Acquisition function combinations tested: LCHC = Low Confidence + High Confidence, HLH = Hybrid Least Confidence + High Confidence, RHC = Random + High Confidence, RLC = Random + Low Confidence.

The hybrid acquisition function consistently achieves the best or near-best accuracy across architectures.

---

DinoV2 on LabelMe1250K — Foundation Model Ablation

Pre-trained DinoV2 outperforms randomly initialized DinoV2 in most settings. However, with the HCLC hybrid sampling strategy, the from-scratch DinoV2 achieves results comparable to the pre-trained version — confirming that the proposed sampling methods are effective even with large foundation models and are readily combined with strong pre-trained backbones in practice.

---

Class Imbalance Ablation (CIFAR-10, 10:1 imbalance ratio, VGG-16)

Method	Accuracy (%)
LC	92.95
HCLC	93.85
HC	92.50

Under class imbalance, HCLC is the preferred strategy. Pure uncertainty sampling (LC) alone can ignore minority classes; the hybrid approach handles skewed distributions better.

---

Repository Structure

├── CIFAR10/
│   ├── densenet121_c10.py          # DenseNet-121 on CIFAR-10
│   ├── resnet18_c10.py             # ResNet-18 on CIFAR-10
│   ├── resnet56_c10.py             # ResNet-56 on CIFAR-10
│   ├── vgg16_c10.py                # VGG-16 on CIFAR-10
│   ├── mobilenet_c10.py            # MobileNetV2 on CIFAR-10
│   ├── swin_c10.py                 # Swin Transformer on CIFAR-10
│   └── smallvit_c10.py             # ViT-Small on CIFAR-10
├── CIFAR100/
│   ├── resnet18_c100_svhn.py       # ResNet-18 on CIFAR-100 / SVHN
│   ├── resnet56_c100_svhn.py       # ResNet-56 on CIFAR-100 / SVHN
│   ├── vgg16_c100_svhn.py          # VGG-16 on CIFAR-100 / SVHN
│   └── mobilenet_c100_new.py       # MobileNetV2 on CIFAR-100
├── SVHN/
│   ├── densenet121_svhn.py         # DenseNet-121 on SVHN
│   ├── resnet18_c100_svhn.py       # ResNet-18 on SVHN
│   ├── resnet56_svhn.py            # ResNet-56 on SVHN
│   └── vgg16_svhn.py               # VGG-16 on SVHN
├── ResNet18 TinyImageNet/
│   ├── resnet18_tin_new.py         # ResNet-18 on TinyImageNet
│   └── glister_TinyImageNet_ResNet18.py  # GLISTER baseline reproduction
├── PascalVOC SSD/
│   └── pascal_voc_ssd.py           # SSD object detection on VOC 2012
├── DinoV2 LabelMe1250K/
│   └── DinoV2_LabelMe1250K.py      # DinoV2 on LabelMe1250K
├── DenseNet121 CIFAR10/
│   └── glister_CIFAR10_DenseNet121.py   # GLISTER baseline reproduction
└── Class Imbalance VGG16/
    └── class_imbalance_vgg.py      # Class imbalance ablation

---

Requirements

pip install torch torchvision numpy scipy matplotlib tqdm
pip install transformers          # DinoV2 experiments
pip install torchmetrics          # Pascal VOC mAP evaluation
pip install timm einops           # Swin Transformer

Tested with Python 3.8+, PyTorch 2.0+. All experiments run on NVIDIA GeForce RTX 4080 (16 GB).

---

Reproducing Results

All scripts share the same structure. Each experiment runs with 3 seeds (42, 789, 101112) and reports mean ± std.

The 4 experiment variants selectable via --exp are:

--exp	Initial selection	Subsequent rounds	Maps to
1	Low Confidence	Low Confidence	LC
2	High Confidence	Low Confidence	HCLC
3	High Confidence	High Confidence	HC
4	Low Confidence	High Confidence	—

---

CIFAR-10

cd CIFAR10
python densenet121_c10.py --all        # runs all 4 variants across 3 seeds
python densenet121_c10.py --exp 1      # LC only
python resnet18_c10.py --all
python resnet56_c10.py --all
python vgg16_c10.py
python mobilenet_c10.py --all
python swin_c10.py --all
python smallvit_c10.py --all

CIFAR-100

cd CIFAR100
python resnet18_c100_svhn.py --dataset cifar100
python resnet56_c100_svhn.py --dataset cifar100
python vgg16_c100_svhn.py --dataset cifar100
python mobilenet_c100_new.py --all

SVHN

cd SVHN
python densenet121_svhn.py --all
python resnet18_c100_svhn.py --dataset svhn
python resnet56_svhn.py --all
python vgg16_svhn.py --all

TinyImageNet

The script will auto-download TinyImageNet from Stanford if not already present (~237 MB).

cd "ResNet18 TinyImageNet"
python resnet18_tin_new.py --all --data-dir ./tiny-imagenet-200
python resnet18_tin_new.py --exp 1 --data-dir ./tiny-imagenet-200   # LC only

Pascal VOC 2012 (Object Detection)

Download Pascal VOC 2012 and place it at ./data/VOCdevkit/VOC2012, then:

cd "PascalVOC SSD"
python pascal_voc_ssd.py              # single seed
python pascal_voc_ssd.py --seeds 3   # 3 seeds for mean ± std

DinoV2 on LabelMe1250K (Foundation Model Ablation)

Set dataset_path in the script to your LabelMe1250K directory, then:

cd "DinoV2 LabelMe1250K"
python DinoV2_LabelMe1250K.py        # runs non-pretrained (random weights)

To also run the pre-trained DinoV2, uncomment the exp1_pretrained_dinov2 call in the __main__ block.

Reproducing GLISTER Baselines

# DenseNet-121 on CIFAR-10
cd "DenseNet121 CIFAR10"
python glister_CIFAR10_DenseNet121.py --num_runs 3 --epochs_per_round 50

# ResNet-18 on TinyImageNet
cd "ResNet18 TinyImageNet"
python glister_TinyImageNet_ResNet18.py --num_runs 3 --epochs_per_round 100

Class Imbalance Ablation

cd "Class Imbalance VGG16"
python class_imbalance_vgg.py

---

How the Pipeline Works

1. Initialize a model with random weights — no pre-training.
2. Select initial samples (~10K or 4% of the dataset) from the unlabeled pool using the chosen confidence criterion applied to the random model's softmax output.
3. Train the model on this labeled set for 100 epochs.
4. Iteratively: select the next batch (~5K or 5%) from the remaining unlabeled data using the trained model's confidence scores, add to the labeled pool, and retrain for 100 epochs.
5. Repeat until the annotation budget is exhausted.

No candidate model is ever trained. The core insight — motivated by the Lottery Ticket Hypothesis — is that randomly initialized networks already produce useful signal for guiding sample selection.

---

Sampling Strategies

Strategy	Acquisition Function	Description
HC	φ_HC(x) = max_k P(y=k|x)	Selects highest-confidence (easy) samples — good for imbalanced data
LC	φ_LC(x) = 1 − max_k P(y=k|x)	Selects most uncertain samples — best for balanced datasets
HCLC	HC first, then LC	Builds stable foundation first, then explores uncertain regions