FD-CAM: Multidisciplinary Methods

• FD-CAM is a suite of methods spanning visual explanation in CNNs, ferroelectric in-memory computing, digital telescope cameras, and numerical solutions for singular differential operators.
• It introduces techniques like grouped channel switching for enhanced model interpretability, sub-nanosecond programmable delays in FeFET arrays, and FPGA-based modular readouts for high-speed imaging.
• Additionally, the numerical FD-CAM method delivers high-precision solutions for Sturm-Liouville problems through superexponential convergence, demonstrating practical efficacy across diverse applications.

FD-CAM (Faithfulness-Discriminability Class Activation Map) refers to a suite of methodologies, algorithms, and hardware concepts across diverse fields such as visual explanation in neural networks, in-memory computing with ferroelectric CAM arrays, high-speed imaging for astrophysical instrumentation, and numerical analysis for singular differential operators. This article provides a rigorous overview of the most prominent incarnations of FD-CAM as described in recent literature, with particular emphasis on the FD-CAM visual explanation method for convolutional neural networks (CNNs), ferroelectric CAM-based in-memory compute macros, Cherenkov telescope digital cameras, and coefficient-approximation methods for Sturm-Liouville problems.

1. Visual Explanation in CNNs: FD-CAM Algorithm (Li et al., 2022)

FD-CAM proposes a class activation mapping scheme that simultaneously enhances faithfulness and discriminability of visual explanations for CNNs. Faithfulness denotes agreement between pixel importance and model prediction changes upon perturbation; discriminability quantifies how well highlighted regions distinguish between different classes.

FD-CAM operates on activation maps $A^k\in\mathbb{R}^{h\times w}$ of a convolutional layer, outputting a class $c$-specific heatmap: $$\mathcal{L}^c_{\mathrm{FD\text{-}CAM}} = \mathrm{ReLU}\left(\sum_{k=1}^K \omega^k_c\,A^k\right)$$ with

$$\omega_c^k = \hat\alpha_c^k\,\exp(\hat s_c^k - 0.5)$$

where $\alpha$ and $s$ are min-max normalized gradient- and score-based channel weights.

The score-based component ($s_c^k$) innovates by introducing grouped channel switching:

• For each channel $A^k$, cosine similarities $M_{kl}$ are computed between all channels, and a similarity group $G(A^k)$ is defined as those with $c$0 above the 5th percentile.
• Both group switch-off (zero $c$1) and group switch-on (retain only $c$2) perturbations are performed, yielding $c$3 and $c$4 respectively.
• The final $c$5 is averaged symmetrically: $c$6.

The gradient-based term $c$7 is

$c$8

This ensures per-class discriminability by leveraging gradients of the target output.

Combination: Empirically, an exponential blend favoring high $c$9 ($$\mathcal{L}^c_{\mathrm{FD\text{-}CAM}} = \mathrm{ReLU}\left(\sum_{k=1}^K \omega^k_c\,A^k\right)$$0) yields the best tradeoff, boosting both faithfulness and discriminability over Grad-CAM, Score-CAM, Ablation-CAM, and recent variants.

Benchmark Results

Method	Insertion$$\mathcal{L}^c_{\mathrm{FD\text{-}CAM}} = \mathrm{ReLU}\left(\sum_{k=1}^K \omega^k_c\,A^k\right)$$1	Deletion$$\mathcal{L}^c_{\mathrm{FD\text{-}CAM}} = \mathrm{ReLU}\left(\sum_{k=1}^K \omega^k_c\,A^k\right)$$2	Overall$$\mathcal{L}^c_{\mathrm{FD\text{-}CAM}} = \mathrm{ReLU}\left(\sum_{k=1}^K \omega^k_c\,A^k\right)$$3	Pointing Game Acc. (%)
Grad-CAM	0.5357	0.1117	0.4240	81.20
Score-CAM	0.5422	0.1059	0.4363	78.46
Ablation-CAM	0.5502	0.1013	0.4489	58.19
FD-CAM	0.5534	0.1001	0.4533	83.70

FD-CAM's grouped channel switching provides notable improvements in both quantitative (insertion/deletion AUC, pointing game) and qualitative evaluations on datasets such as ILSVRC2015 and PASCAL VOC 2007 (Li et al., 2022).

2. Ferroelectric CAM (FeFET): FD-CAM in Time-Domain In-Memory Computing (Mattar et al., 4 Apr 2025)

A separate usage of FD-CAM refers to FeFET-based Content-Addressable Memory leveraged for time-domain nonvolatile in-memory computing (TD-nvIMC). This architecture achieves energy efficiency by integrating a FeFET CAM array with delay element chains and on-die time-to-digital conversion (TDC) in 28 nm CMOS technology.

Key architectural features:

• Each CAM cell comprises complementary FeFETs (F+, F–) programmed to high/low threshold states (HVT/LVT).
• Matching is sensed via discharge delay ($$\mathcal{L}^c_{\mathrm{FD\text{-}CAM}} = \mathrm{ReLU}\left(\sum_{k=1}^K \omega^k_c\,A^k\right)$$4 for match, $$\mathcal{L}^c_{\mathrm{FD\text{-}CAM}} = \mathrm{ReLU}\left(\sum_{k=1}^K \omega^k_c\,A^k\right)$$5 for mismatch) at a current-starved inverter, with row-wise select-line pulsing and word-line control.
• TDC block digitizes the integrated delay per row, encoding binary MAC or Boolean logic (XOR/AND).
• Multilevel delay calibration corrects for device/path mismatch, achieving 100 ps resolution (σ_post ≈ 30 ps), with $$\mathcal{L}^c_{\mathrm{FD\text{-}CAM}} = \mathrm{ReLU}\left(\sum_{k=1}^K \omega^k_c\,A^k\right)$$60.15× variance reduction compared to uncalibrated operation.

Performance metrics:

Parameter	Value
Technology	28 nm CMOS FeFET
Delay step (t_dH–t_dL)	550 ps
Calibration resolution	100 ps
Throughput	232 GOPS
Energy efficiency	1887 TOPS/W
Supply voltage	0.85 V
AND-MAC latency (3 bits)	≈550 ps
XOR-MAC latency (3 bits)	≈1.3 ns

This integration of FD-CAM cells with isolated bulks, multi-level delay calibration, and robust write-disturb prevention realizes a nearly 2000× improvement in programmable delay step and system-level energy efficiency over prior TD-nvIMC systems (Mattar et al., 4 Apr 2025).

3. Fully-Digital Camera Systems: FD-CAM in Atmospheric Cherenkov Telescopes (Pühlhofer et al., 2015)

FD-CAM also denotes the “FlashCam” fully-digital camera system for the Cherenkov Telescope Array’s (CTA) medium-sized telescopes. Here, FD-CAM designates the camera’s modular digital readout chain, with the following architectural divisions:

• Photon Detector Plane (PDP): Hexagonal PMT matrix, preamplifiers/slow control.
• Front-End Readout System (ROS): Commercial 12-bit FADCs @250 MS/s, Spartan-6 FPGAs, real-time buffering, cluster triggering, and per-pixel digital processing.
• Camera Server: 10 GbE data links, high-throughput DAQ, dead-time-free streaming (>2 GB/s, >30 kHz sustained event rates).

Key metrics:

• Signal digitization at 4 ns samples, software interpolation to ≲2 ns resolution.
• Dynamic range: 1–1000 p.e. linear, up to 5000 p.e. non-linear, with <5% charge error above 10 p.e.
• FPGA-based local clustering, per-pixel calibration, and maintenance-friendly hardware modularity (Pühlhofer et al., 2015).

4. Numerical Solution of Singular Differential Operators: FD-CAM Method (Makarov et al., 2011)

In numerical analysis, FD-CAM refers to the Functional-Discrete Coefficient Approximation + Homotopy Method for singular Sturm-Liouville eigenvalue problems. The method proceeds as follows:

• The coefficient approximation method (CAM) replaces $$\mathcal{L}^c_{\mathrm{FD\text{-}CAM}} = \mathrm{ReLU}\left(\sum_{k=1}^K \omega^k_c\,A^k\right)$$7 by a piecewise-constant $$\mathcal{L}^c_{\mathrm{FD\text{-}CAM}} = \mathrm{ReLU}\left(\sum_{k=1}^K \omega^k_c\,A^k\right)$$8, enabling analytic base problem solutions.
• Homotopy connects the base to the full problem, expanding eigenvalues/functions in a formal series in $$\mathcal{L}^c_{\mathrm{FD\text{-}CAM}} = \mathrm{ReLU}\left(\sum_{k=1}^K \omega^k_c\,A^k\right)$$9.
• Correction terms $$\omega_c^k = \hat\alpha_c^k\,\exp(\hat s_c^k - 0.5)$$0, $$\omega_c^k = \hat\alpha_c^k\,\exp(\hat s_c^k - 0.5)$$1 are computed recursively with orthogonality conditions, typically truncated at finite $$\omega_c^k = \hat\alpha_c^k\,\exp(\hat s_c^k - 0.5)$$2.
• The series exhibits superexponential convergence ($$\omega_c^k = \hat\alpha_c^k\,\exp(\hat s_c^k - 0.5)$$3, $$\omega_c^k = \hat\alpha_c^k\,\exp(\hat s_c^k - 0.5)$$4), allowing high-precision solutions with modest $$\omega_c^k = \hat\alpha_c^k\,\exp(\hat s_c^k - 0.5)$$5, particularly for large eigen-indices and singular coefficients.

Typical convergence is demonstrated by errors decreasing from $$\omega_c^k = \hat\alpha_c^k\,\exp(\hat s_c^k - 0.5)$$6 to $$\omega_c^k = \hat\alpha_c^k\,\exp(\hat s_c^k - 0.5)$$7 by $$\omega_c^k = \hat\alpha_c^k\,\exp(\hat s_c^k - 0.5)$$8 in Legendre-type examples (Makarov et al., 2011). The method is robust to singular integrands and is competitive with SLEIGN2.

5. Comparative Synthesis and Field-Specific Distinctions

The designation FD-CAM refers to distinct concepts across several research sectors:

Context	FD-CAM Meaning	Primary Attributes
CNN Visual Explanation	Faithfulness-Discriminability CAM	Grouped channel switching, hybrid gradient/score weighting (Li et al., 2022)
In-Memory Computing	Ferroelectric FeFET CAM for TD-nvIMC	Delay-encoded MAC, sub-nanosecond programmable delay (Mattar et al., 4 Apr 2025)
Astrophysics Instrumentation	FlashCam fully-digital Cherenkov camera system	Modular digital PMT readout, FPGA DAQ (Pühlhofer et al., 2015)
Numerical Analysis	Functional-Discrete Coefficient Approx. + Homotopy	Superexponential S-L eigen solver (Makarov et al., 2011)

Each usage is domain-specific; the unifying thread is “CAM” (Content-Addressable/Activation Map/Approximation Method) augmented by a leading innovation (“F” for Faithfulness/Ferroelectric/Flash/Functional-Discrete).

6. Key Implementation Insights and Impact

• The FD-CAM visual explanation method is currently the state-of-the-art technique for balancing class specificity and attribution reliability in CNN heatmaps. Both grouped perturbations and exponential fusion are empirically validated to improve class localization, even in multi-instance scenarios (Li et al., 2022).
• FD-CAM (FeFET) architectures significantly advance time-domain in-memory compute macros, reducing step size and improving integration for binary MAC operations in near-memory logic (Mattar et al., 4 Apr 2025).
• The FD-CAM (FlashCam) delivers real-time, dead-time-free digital readout for ground-based gamma-ray astronomy, meeting strict dynamic range, noise, and throughput criteria for the CTA (Pühlhofer et al., 2015).
• In numerical computation, FD-CAM achieves high-accuracy solutions for singular Sturm-Liouville problems at practical truncation orders, with theoretical convergence guarantees even for large eigenindices and singular potentials (Makarov et al., 2011).

Collectively, FD-CAM nomenclature encapsulates cutting-edge advances that combine modularity, hybridization of techniques, and context-aware calibration or perturbation to solve central challenges in interpretability, hardware efficiency, sensor design, and mathematical modeling.